Website Tour

Okay, let me walk you through my website, Tools in Vision, and explain what each part does. The part I want to spend the most time on is AnnoTrack. That's where we label images, work with video frames, track objects, review annotations, and prepare datasets for training. Around AnnoTrack, there are several other tools. AI Studio creates images, videos, synthetic datasets, and narration. Media Tools processes existing images and videos and checks datasets. The professional annotation service lets someone request labeling work. And the Learning Center explains the ideas behind these tools. I'll use one example throughout the walkthrough. Let's say we're working on a project that needs to identify cars in a parking lot. We start with a video. We extract useful frames from it, label the cars, check the labels, and export the dataset. If several people are working together, we also need a way to review their work and see what's finished. That example gives us a reason for each part of the website. The media tools prepare the files. AnnoTrack creates the labels. The team features organize the review. And the Learning Center explains concepts like bounding boxes, segmentation, and training data. The prices and limits I'll mention belong to this version of the website. The account and tool pages show the settings available to the user. Let's start with the home page. The main navigation takes us to AI Studio, AnnoTrack, Annotation Service, Media Tools, and Learning Center. These areas use the same Tools in Vision account. The user doesn't need a separate account for every tool. The account page gives a quick overview of the profile, AI credit balance, and AnnoTrack subscription. The subscription and the AI credit balance serve different purposes. The AnnoTrack subscription controls things like storage, export access, project AI training, and team features. AI credits are a prepaid balance used for generation. So buying an annotation subscription doesn't mean every AI generation is included. The account page shows those separately so the user can understand what they have. The sign-in page supports email and password. It also has Google and GitHub sign-in options when those providers are configured. There is email verification, password recovery, and an account security page where the user can check recognized devices and login history. Once I'm signed in, I can open AnnoTrack and choose a project. A project keeps a dataset organized. I could have one project for parking-lot cars, another for pedestrians, and another for an industrial inspection task. The Personal Workspace contains my own projects. The Team Workspace contains projects shared with the team. Each workspace has its own access and project list. Being a viewer in a team doesn't remove the user's ability to work in their own personal workspace under their personal plan. On the project page, I can create a project, open it, rename it, or delete it. There are also copy and move controls between the personal and team workspaces, when the user's permissions allow that. Copy keeps the project in the original workspace and creates a copy in the destination. Move transfers the project to the other workspace. Once I open a project, the image appears on the left. This is the canvas where I draw and edit annotations. The controls are on the right. They show which account, workspace, and project I'm using. They also show storage information, the image name, the annotation mode, and the available tools. The project information becomes especially useful when we get to AI tracking. Each project has its own training examples and its own trained model. A model trained for one project doesn't automatically become the model for every other project. For our parking-lot example, the next step is adding images. If the images are already prepared, I can select several files and upload them together. The page shows upload progress and the limits for the account. The actual limits depend on the plan and the available storage. If I have a video, I can extract its frames directly into the project. I choose the video, select the part I want, and decide how often to extract a frame. There are two ways to describe the interval. I can use time, such as one image every second. Or I can use a frame interval, such as every thirty frames. Those mean the same thing only when the source has the appropriate constant frame rate. With a variable frame rate, the difference matters. I can choose JPEG, PNG, or WebP for the extracted images and set a maximum number of frames. The page gives an estimate before extraction. After extraction, each frame becomes a normal image in the annotation project. For tracking, I want those images in the correct sequence. If I jump from one unrelated picture to another, there isn't a sensible movement for the tracker to follow. If I sample too far apart, the object might move a long distance between images, which can make tracking more difficult. There's also a separate frame extractor in Media Tools. The difference is where the frames go. The video import inside AnnoTrack puts the frames directly into the project so I can annotate them immediately. The separate extractor gives me a ZIP file of frames that I can download and use elsewhere. Once the images are in the project, I can use Previous and Next or choose a particular image from the image list. I don't have to open every image in order when I'm checking a specific frame. Now let's talk about bounding boxes. A bounding box is a rectangle around an object. It tells a detection model where the object is and which class it belongs to. For this example, I select the car class and draw a rectangle around a car. I can select that box, move it, and adjust its edges until it fits the object properly. There's also an Add New Box In Center button. That creates a starting box that I can move and resize. The class is the name of the thing we're labeling. It could be car, truck, person, or another category defined for the project. The class controls let me add a class, rename a selected class, delete a class, or assign the selected class to an existing box. Renaming a class is handled across the saved project annotations so its name stays consistent. The box and the class describe two different things. The coordinates describe the location. The class describes what is at that location. A perfectly placed box with the wrong class is still a bad label. If I make a mistake, I can delete the selected box. There is also a button to delete all boxes on the current image. That is separate from deleting the image itself. The image deletion controls are in the Images section. For a large image, I can zoom in with the mouse wheel. I can pan by holding the Space key while dragging. Fit Image and Reset View bring me back to the full view. The minimap helps me see which part of the image I'm looking at while zoomed in. There are keyboard shortcuts too. The A key or the left arrow goes to the previous image. The D key or the right arrow goes to the next image. Delete removes the selected annotation. Control and Z handle undo. Control and Y handle redo. The workspace has autosave and a save-status indicator. I still watch that indicator and use the explicit Save button when I want to confirm that the work is saved. Tracking suggestions have an extra review step. I'll explain that after we look at segmentation.

Now let's move to segmentation. With a bounding box, I get a rectangle around an object. With a polygon, I can follow the object's outline more closely. For example, a car viewed at an angle has an outline that doesn't match a simple rectangle. I click around that outline to place the polygon points. I can finish using the Finish Polygon button, the Enter key, or a double-click. Escape cancels the polygon I'm drawing or clears the selection. After finishing the polygon, I can adjust its points. If I need more detail on an edge, I can insert another vertex. A vertex is simply one of the points that defines the polygon. If a point is unnecessary, I can delete it. Smooth Selected Polygon adjusts the contour. Simplify Selected Polygon reduces the number of points. After either operation, I would inspect the result, especially around narrow or detailed parts of the object. There are three useful polygon properties: Occluded, Hidden, and Locked. Occluded marks an object as partially blocked. Hidden hides the polygon from the screen so I can see other objects more clearly. Locked protects the polygon from accidental editing. There are also controls to unhide or unlock all polygons when needed. Occluded is an annotation property. Checking it doesn't reconstruct the hidden part of the object. I still need a clear labeling rule for the boundary that the project expects me to draw. There are two ways to carry polygon work forward. Carry selected polygon across images copies the selected shape to the next opened image as a starting point. Polygon Tracking uses image movement to estimate where the polygons should go in the next frame. When Polygon Tracking is enabled, pressing Next attempts to track each polygon independently. Existing destination annotations are preserved. The polygon tracker uses optical flow and feature-based motion estimates. In simple terms, it looks at how visible patterns move between the images and uses that movement to propose a new polygon position. I still need to inspect the shape and correct it when the estimate is off. Consecutive images need matching dimensions for this polygon-tracking workflow. If a car moves slightly to the right, the proposed polygon can move with it. If the car turns, becomes blocked, or changes its visible outline, I may need to adjust the points more carefully. The polygon tracker is separate from the project-trained YOLO box tracker. The project AI training controls use saved bounding boxes. They don't train a segmentation model from these polygons. Now let's return to bounding boxes and look at the four tracking modes. The first is Manual Mode. Manual Mode means tracking is off. I draw and edit everything myself. It's useful when the images are unrelated or when I want complete manual control. The second is Simple Tracking. Simple Tracking follows the appearance and movement of the selected object between images. It uses the available OpenCV tracking methods, with motion and template matching as fallbacks. It doesn't require a trained project model. The third is Fast Tracking. Fast Tracking uses the project's trained YOLO detector with ByteTrack. The detector finds objects in the images. The tracker then tries to match the selected object between consecutive frames. The fourth is Stable Tracking. Stable Tracking uses the same project's detector with BoT-SORT. It's another tracking option that I can compare on my footage. I would judge the result on the actual images instead of assuming that a mode will always perform better because of its name. The AI modes require Pro Plus or Team access. The particular project also needs a successfully trained model. Until those conditions are met, the AI modes are unavailable. That explains why they might be enabled in one project and unavailable in another. I can switch modes with the buttons or with keyboard shortcuts. Control, Alt, and zero turn tracking off. Control, Alt, and one select Simple Tracking. Control, Alt, and two select Fast Tracking. Control, Alt, and three select Stable Tracking. Those shortcuts choose the mode. To move the selected box to the next image, I then use Next, the D key, the right arrow, or the tracking button. The current mode is displayed in the interface. The choice is remembered for the relevant project, but the page still checks whether that mode is available when I return. Now let's look at how a project gets its own AI model. The Project AI Model panel starts with the saved annotated image count. The current training range is one hundred to one thousand saved images with valid labeled bounding boxes. This is a count of images, not a count of objects. If I draw ten boxes on one image, that gives me one training image. If I upload five hundred pictures without saving valid box labels, uploading them alone doesn't give me five hundred training examples. If the project goes past one thousand eligible images, I can still annotate it, but I can't start a new training run in this version. The trainer doesn't silently select just the first thousand images from a larger project. The training count is separate from the formal Team approval process. It comes from the saved valid box annotations available to the training workflow. Once the project is in the supported range, and the account has the required access, I can click Train AI Model. Saving more images doesn't automatically start another training run. The user chooses when to train or retrain. Training starts from a pretrained YOLO eleven N model and adapts it to the labels in this project. The code creates a snapshot of the saved labels for the run, prepares the images, and separates them into training and validation sets. For one hundred eligible images, the current split gives eighty training images and twenty validation images. The preparation code puts every fifth image into validation. So twenty out of one hundred are held out. This is a basic validation split. If the pictures are neighboring frames from one recording, I would still need a separate evaluation on different recordings before claiming that the model generalizes. Also, one hundred images is the minimum that enables training. It doesn't guarantee that the examples cover the task well. Useful examples should include variation in viewpoints, lighting, object appearance, and difficult cases. Repeated almost-identical frames can meet the count without adding much useful variety. The default training configuration uses thirty epochs, although the server configuration can change that. An epoch is one pass through the training dataset. The trainer can also stop early if its early-stopping condition is reached. Training runs in a background worker. That means I can keep using the annotation workspace while the job is waiting or training. The queue runs one training job at a time. The panel shows the stage, queue position when waiting, progress, epoch, image counts, elapsed time, and the model version associated with the run. The progress percentage comes from preparation and training work being completed. It tells me how far the job has reached. It doesn't tell me the model's accuracy. At the end, the system measures detection metrics on the validation images and checks that the trained model can run inference. A completed run means a usable model was produced and activated. It doesn't mean the model is perfect or that it passed a guaranteed accuracy threshold. The panel also shows how many images the current model used, how many new examples have been saved since then, and whether existing examples were corrected. The interface suggests retraining after roughly fifty to one hundred additional corrected and varied images. That's a practical reminder. It doesn't promise a particular improvement. If I cancel an unfinished run, its unfinished training progress is discarded. My annotations remain available, and the previous successful model stays active. The same protection applies if retraining fails. A failed new run doesn't replace a working model. Once a model is available, I can use it to track an object. I select the car's box, choose the tracking mode, and advance to the next image. The tool estimates where that selected car should be in the next frame. I can inspect the proposed box, move it, resize it, or correct its class before saving. The AI proposal stays as an editable draft until I save it. That gives me a chance to catch a drift or a wrong match before using the proposal as a saved label or another training example. If the detector can't confidently match the selected object, the implementation can fall back to a local motion estimate or retain the previous position for correction. So seeing a box on the screen doesn't prove that the AI recognized the object correctly. I still review the draft. The workflow also protects destination work. If a destination already has relevant saved annotations, the system checks them and avoids blindly replacing the work. A save conflict needs to be resolved instead of silently overwriting a newer version. For a difficult frame, I can switch back to Simple Tracking or Manual Mode, fix the box, and save it. Later, I can include those corrected examples in a new training run. There is also a button called Propagate Current State Forward. That copies the current boxes and class state to following images until it reaches an image marked as explicitly annotated. It doesn't estimate the object's movement. It can be useful for repeated starting positions. But if the car moves, I'll need to edit the copies. For actual movement, I use the tracking action and inspect each next frame.

Now let's look at how AnnoTrack works when several people share a project. Different team members can have different responsibilities. The owner manages the team and its billing. An admin can help manage members, projects, and seats without getting the owner's billing permission. An annotator creates and saves labels. A reviewer handles submitted work and approves or rejects it. A quality assurance lead, or QA lead, can also manage quality-check samples. A viewer can inspect the team project but can't edit annotations or export the dataset. Reviewers and QA leads also have annotation-editing permission in the current role setup. The interface controls and server permissions work together to limit what each role can do. To invite someone, the team manager enters the person's email, chooses an available role, and sends an invitation. The member can accept that invitation from their account. The team page shows current members and pending invitations. It also has controls for managing paid seats. This supports a workflow where one person labels images, another reviews them, and a third person checks a sample of the approved work. Everyone can work from the same project instead of passing around several different copies of the dataset. Once an annotator is happy with an image, they can submit it for review. If several images are ready, the annotator can select those images and submit them together. Only the selected images enter that submission. For example, I can submit twenty finished images while continuing to work on the rest of the project. The review queue groups the submission so the reviewer can see the waiting work. The reviewer can filter the queue to the current project or look across the team's projects. Completed submissions are available too. When a reviewer opens a submission, the workspace enters Reviewer Mode. The review toolbar lets them move between submitted images, add a review comment, and approve or reject the current item. Let's say I labeled a car, but the box cuts off its bumper. The reviewer can reject the annotation and explain what needs changing. I can read the feedback, correct the box, save it, and submit the image again. The annotation status can be Draft, In Review, Approved, or Rejected. Bounding-box annotations and segmentation annotations are tracked by their annotation type. That lets me distinguish a review of the boxes from a review of the polygons. An approved annotation has been accepted through the review workflow. Its quality still depends on the labeling rules and the care of the reviewer. Approval isn't an automatic measurement of ground-truth accuracy. The review history records decisions and feedback. That helps us understand what happened to the work without relying on someone remembering which files they checked. After review, the QA tools let the team check a random sample of approved annotation records. The person managing QA chooses a percentage and a team member to handle the sample. For example, if one hundred approved records are eligible for a new sample, choosing ten percent gives ten QA items. That's one hundred times ten, divided by one hundred. The sampler uses approved records that don't already have a QA assignment. The QA result can be passed or failed. It can include a score from zero to one hundred, a comment, and an error category when needed. So I can record whether the problem was a wrong class, a poor boundary, or another labeling issue. Those scores are entered by the person doing the check. They are review scores. A neural network isn't automatically calculating whether every annotation is correct. The team dashboard brings the work together. It shows image and project counts, team seats, storage, waiting reviews, approved and rejected records, approval rates, review time, and QA progress. The contributor table shows submitted, approved, and rejected work, reviewer workload, and available QA scores. If a person hasn't been rated, the interface shows that instead of inventing a score. I can use the dashboard to see whether a lot of labeling is finished but waiting for review, or whether certain work needs another look. The numbers help manage the workflow. I still open the images when judging the quality itself. There are also comments and communication tools. Comments stay connected to the image we're viewing. That means I can discuss the actual frame instead of saying, the third picture in the folder, and hoping someone opens the same one. A comment can also target an annotation object. For example, I can point out that a particular car box is too wide. Replies keep the discussion together, and a comment can be resolved when the issue has been dealt with. Mentions bring a teammate's attention to a comment. The comment remains visible to team members who can view that image. The mention notification goes to the person I mention. The Activity tab shows recorded events. I can filter by person, event type, or date. That helps when I want to understand when an annotation changed or when a review was completed. For a broader conversation, there is project team chat. There are also private messages between participants. Notifications collect things like mentions, review updates, QA assignments, and ownership requests. Presence shows active or idle collaborators. When people work on the same image, live collaboration can show cursor positions and annotation updates. That's useful when I'm explaining a labeling decision and want the other person to see which object I'm pointing at. The collaboration window also has a Help tab. It explains the roles and the review workflow, and there is a downloadable team collaboration guide. Now, what happens if two people change the same annotation? The conflict dialog offers four choices. Use the remote version. Keep my local version. Merge both versions. Or request ownership. The remote version is the version saved elsewhere. The local version is the one I'm working with in my browser. Merge attempts to bring the versions together. I still review the merged result because two boxes could describe the same object differently. If one person needs to handle the next save, I can request ownership from the relevant teammate. That person can grant or decline the request. A grant is for one use and expires after sixty minutes. This gives the users an explicit decision instead of quietly losing someone's edits. It also means collaboration involves handling who can save, which version they're saving, and how competing changes get resolved. Once the labels are ready, I can check the dataset and export it. Validate Dataset reports the number of images, annotated images, empty images, boxes, and polygons. It also shows class counts and lists issues and warnings. For bounding boxes, it can flag invalid rectangle data, coordinates outside the image, and missing class references. For polygons, it checks point structure and related class information. It can also warn about duplicate class names. These checks catch structural problems. I still need visual review to know whether I labeled the correct car or missed another car in the scene. A file can have valid coordinates and still describe the wrong object. After checking, I choose the export formats I need. For bounding boxes, the options include JSON, YOLO text, COCO JSON, Pascal VOC XML, CSV, and images with the annotations drawn on them. Export access depends on the plan. JSON keeps structured annotation data. CSV is useful when I want to inspect the records in a table. Annotated images let a person quickly see how the boxes look. YOLO, COCO, and Pascal VOC are formats that different training and evaluation tools expect. For a YOLO bounding box, one row contains a class ID and four normalized values. Those values are the center position across the image, the center position down the image, the box width, and the box height. For example, class zero followed by point five, point five, point two, and point four describes a box centered halfway across and down. Its width is twenty percent of the image width, and its height is forty percent of the image height. The horizontal measurements are normalized by dividing them by image width. The vertical measurements are normalized by dividing them by image height. For segmentation, the exports include polygon JSON, YOLO segmentation text, COCO segmentation JSON, CSV, annotated previews, and PNG masks. YOLO segmentation stores the polygon's normalized point coordinates. Its rows are different from the five-value bounding-box rows. A mask turns polygon regions into pixel labels. An annotated preview draws the outlines on an image for viewing. I choose a mask when the workflow needs a pixel-label image, and a preview when I want to inspect the annotation visually. Save Current As handles the current image's export. Create Dataset ZIP packages the project in the selected formats. ZIP creation runs as a job with progress, a cancel control, and a download link when the result is ready.

Let's look at the plans and storage choices. This version lists Free, Pro, Pro Plus, and Team. Free includes five hundred megabytes of managed storage and a ten-megabyte maximum image size. The current code allows JSON exports on the Free plan. Pro is listed at nine dollars and ninety-nine cents a month. It includes fifteen gigabytes of storage and a fifty-megabyte maximum image size. It also adds the other supported export formats. Pro Plus is listed at fourteen dollars and ninety-nine cents a month. It includes one hundred gigabytes of storage and a one-hundred-megabyte maximum image size. Pro Plus and Team enable the project AI training and AI bounding-box tracking features. Team is listed at thirty-nine dollars and ninety-nine cents a month. It includes two hundred and fifty gigabytes of shared storage and three users. Additional users are listed at four dollars and ninety-nine cents each per month. For example, one extra user makes the listed base total forty-four dollars and ninety-eight cents before other billing adjustments. That's thirty-nine dollars and ninety-nine cents, plus four dollars and ninety-nine cents. Users can manage their subscription, review billing information, update payment details, and cancel through the subscription-management controls. For Pro Plus and Team, there is also an optional storage feature called bring your own storage. The user can connect an S3-compatible bucket, including Amazon S3, Cloudflare R2, Backblaze B2, or Wasabi. They save the connection, test it, and choose whether new uploads should use that bucket. Managed storage is already included in the plan, so connecting a separate bucket is optional. When enabled, the new uploads use the user's bucket instead of the included managed storage. The storage provider's own costs and connection settings still belong to that separate storage setup. Now let's move to AI Studio. The general workflow is to describe the result, choose the settings, check the credit price, and create the job. The recent-creations section shows progress and provides the output when it's ready. The Image Generator creates images from a written description. For example, I could describe a parking lot, the camera viewpoint, the lighting, the weather, and the objects that should be visible. I can choose the number of outputs, quality, and square, landscape, or portrait shape. The supported request range is one to ten images. The listed quality choices cost two credits per draft image, eight per standard image, and twenty-nine per premium image. So four standard images cost thirty-two credits. That's four times eight. The price updates before creation. The Image Editor starts with uploaded images. I can provide one to ten reference images and describe the change I want. Then I choose the number and settings of the output images. This is an AI-directed edit, so I inspect the result to check whether the parts I care about were preserved. If I need to resize, crop, or convert a file using exact settings, I can use the Image Toolkit in Media Tools. The visual-generation code sends requests through a configured Runway API connection. The website handles the account, settings, job progress, charging, and private output delivery around that generation service. The Video Generator starts with a text description and can also use a reference image. I choose the duration and resolution, then check the displayed price. The duration choices are five, ten, fifteen, thirty, and sixty seconds. Longer outputs are made from shorter generated scenes and combined into one MP4 file. The scene prompts ask for the same subjects, setting, and visual style throughout. I would still inspect the scene joins and subject consistency. Asking a model to preserve a character doesn't guarantee that every scene will match perfectly. The listed resolutions are four eighty P, seven twenty P, and ten eighty P. When a reference image is used, the current video-generation workflow supports up to seven twenty P. To use ten eighty P, the request needs to be text-only. The price depends on duration and resolution. In this version, seven twenty P costs eighteen credits per second. A five-second request at that resolution costs ninety credits. That's five times eighteen. The interface calculates the selected request before submission. Generation can take time, so for a walkthrough I can explain the settings and show an existing result instead of waiting for a new video to finish. AI Studio also has a tool called Create Your Dataset. This creates a collection of generated images or frames. The maximum requested size is three thousand files. There are two modes: independent images and sequential frames. Independent images are generated as separate image outputs from the written prompt. I choose the count, image quality, and shape. In this implementation, the independent-image workflow uses the prompt. It doesn't apply the optional reference image as guidance. Sequential frames work differently. The tool generates video clips and extracts frames at the selected rate. The choices are six, twelve, or twenty-four frames per second. I can also choose the resolution. The sequential workflow can use a reference image, with the same seven twenty P restriction for image guidance. Sequential output contains neighboring frames within the generated clips. If a large request needs more than one clip, I would inspect the boundaries between clips. The files aren't proof of one physically consistent continuous recording. The results come in ZIP parts with up to one hundred files per part. Each part includes a manifest. The manifest records information such as the prompt, dataset mode, and file range. That makes a large result easier to download and identify. The output is synthetic imagery. It doesn't include automatically reviewed object labels. If I need a labeled detection dataset, I still need to annotate the images, check the labels, and evaluate whether the generated examples are useful for the real task. I would also keep synthetic examples identifiable when studying how they affect model performance.

Now let's look at Voice Studio. Voice Studio turns written text into narration. There are two engine choices on the page. Each needs its corresponding processing service to be configured and enabled for the account. The controls show whether that option is available. Tools in Vision Voice Studio uses the hosted voice service for English narration. It has nineteen American and British preset voices. It also supports cloning from a permitted voice recording. The listed price is ten credits for each started block of five hundred characters. Premium Voice Studio uses ElevenLabs for preset voices and permitted cloning. Its listed price is thirty credits for each started block of five hundred characters. The premium controls include language choices such as English, Arabic, Spanish, French, German, and Turkish. I can play a free voice sample before choosing. Then I paste the narration text, select a preset or clone mode, choose relaxed, natural, or faster speed, and select MP3 or WAV. The text limit is ten thousand characters per request. For a clone, I upload a clean recording of one speaker and confirm that I own the voice or have permission to use it. A recording with music, background noise, or several people talking can make the reference less useful. The hosted workflow removes the temporary reference after processing or cancellation. The premium workflow requests removal of its temporary provider clone. The pricing rounds up by started blocks. For example, six hundred characters use two blocks of five hundred characters. That costs twenty credits in Tools in Vision Voice Studio or sixty credits in Premium Voice Studio. The request price appears before I start. Narration jobs are queued. The page shows the queue position and an approximate completion time. That estimate can change with the queue and service availability. I can leave the page and return later. The hosted option can take longer, so I would choose based on the result I need and the processing capacity available. The output is one audio download, even when the text needs internal processing in several pieces. The audio workflow also balances the volume before delivery. The AI tools share one prepaid balance. The current packages are five hundred credits for five dollars, one thousand and fifty credits for ten dollars, and two thousand seven hundred and fifty credits for twenty-five dollars. The larger packages include extra credits. The website shows the request price before generation and deducts the charge when the job is accepted. Failed paid generation jobs have an automatic credit-return workflow. The credits are described as not expiring. Private output downloads are available for twenty-four hours, so I download the files I want to keep. Stripe or PayPal can be used for the one-time credit purchase when the payment option is available. These purchases are separate from the AnnoTrack subscription. There is also provider billing behind the scenes. Runway and ElevenLabs have their own balances or usage allowances. A customer's Tools in Vision credits aren't the same units as the provider's credits or narration allowance. The owner capacity panel reads the information those providers make available. For example, it can read a Runway balance. That balance response doesn't prove that automatic billing is enabled. The ElevenLabs check may show a subscription allowance without being able to verify a pay-as-you-go balance. For those details, I check the provider account. The capacity panel is a status check. It doesn't purchase provider credits automatically. If a service can't accept more work, the job workflow can show a capacity wait or a failure instead of pretending that an output is ready.

Let's move to Media Tools. These tools start with files I already have. The Image Toolkit lets me process one image or a batch using the same selected settings. I can crop, resize, compress, convert the format, rotate, or flip the images. The resize choices include fitting inside a size, filling and cropping, fitting with padding, or using exact dimensions. There is also an option to allow upscaling. For appearance changes, I can adjust brightness, contrast, saturation, sharpness, blur, grayscale, and auto contrast. I can add text, choose its size and color, and add a prefix or suffix to the batch filenames. In this version, added text uses a built-in font and appears at the top left. The output choices include WebP, JPEG, PNG, GIF, BMP, and TIFF. If I convert a transparent image to JPEG, it needs a background color because the JPEG result doesn't retain alpha transparency. The interface lets me choose that background color. The image page shows the original preview and the processed result. I can also choose to remove metadata, including location information. Then I process the files and download the results from Media Jobs. The Video Toolkit gives me a similar workflow for video. I upload a file and choose the action I want. I can trim a time range, crop the frame, resize, compress, or convert to MP4, WebM, MOV, AVI, or MKV. For audio, I can keep it, remove it, or extract it as an MP3 file. I can also make a GIF or extract frames to a ZIP. The adjustment controls include rotation, horizontal or vertical flipping, speed changes, grayscale, frame rate, and metadata removal. The processing summary helps me see which changes I've selected before running the job. Video compression has a quality tradeoff. A smaller file can lose detail, and the final size can differ from an estimate. The job list shows progress and downloads. These media jobs are kept separate from annotation projects and professional service orders. The current allowance is three free guest jobs and seven free jobs for a free signed-in account. Paid plans remove that job-count quota. Upload, output, duration, and other processing limits still apply. The page shows the current limits so users don't have to guess from a plan name. There are also eight focused dataset and vision utilities. Each has a guide explaining the expected input, the output, and the limitations. The first is Dataset Health Check. This looks at the file structure, whether images can be read, their dimensions, and whether image and label names match. It can flag an image with no label or a label with no corresponding image. It produces a report without changing the original ZIP. The second is Annotation Format Converter. It takes a YOLO bounding-box dataset and converts it to COCO JSON or Pascal VOC. The current converter expects the five-value box format. It doesn't convert YOLO segmentation polygons or every possible annotation format in both directions. The third is Video Frame Extractor. I choose a time range, an interval in seconds or frames, an output image format, and a maximum count. The tool gives me a ZIP of extracted images. For example, a twenty-second range sampled every two seconds gives about ten frames. That's twenty divided by two. This is a planning estimate. The decoded output is what I actually inspect. The fourth is Dataset Splitter. It makes training, validation, and test folders with matching labels. Its defaults are seventy, twenty, and ten percent. It has a shuffle seed so the same input can be split repeatably. With one hundred images, those percentages request seventy training images, twenty validation images, and ten test images. The splitter works at the file level. It doesn't group by video or recording session and doesn't stratify by class. If my data comes from several videos, I would organize the split by recording when I need to test performance on unseen videos. Randomly separating neighboring frames can give an overly easy test. The fifth utility is Duplicate Image Finder. It reports exact duplicates and images with matching compact visual hashes. It doesn't delete anything. I inspect a visually similar group before deciding what to remove. The visual hash can group different pictures or miss some near-duplicates. The sixth is Annotation Quality Checker. This checks supported YOLO box rows for malformed lines, invalid value ranges, and empty or missing labels. It helps with formatting. It doesn't know whether a box encloses the intended object. It also isn't a complete geometric or segmentation-label audit. The seventh is Auto-Anonymize. It detects frontal faces in a still image and blurs the detected regions. I can choose the blur strength. Then I inspect the output because the detector can miss a profile, a small face, or a blocked face. It doesn't automatically blur license plates, identifying text, or an entire video. The eighth is Background Remover. It estimates the main foreground and exports a transparent PNG. I would inspect the edges over light and dark backgrounds, especially around hair or low-contrast boundaries. The foreground estimate can help prepare an image. It isn't automatically a reviewed segmentation label. The dataset guides also offer a small example ZIP. It contains simple labeled rectangles and a duplicate example. That lets someone learn the file workflow before uploading their own dataset. Now let's look at the professional annotation service. This is for someone who wants annotation work done as a service. It's a separate one-time order. The user enters project information, requested classes, annotation instructions, approximate image count, preferred export format, and requested completion date. They can provide a private cloud link or upload a dataset ZIP directly. There are also places for a sample image and an annotation example. Submitting the quote request doesn't collect payment. The service first inspects a representative sample, around ten to twenty images, to understand the actual scope and prepare the quote. The listed starting prices are six cents per bounding-box object and eighteen cents per standard polygon object. Standard polygons are described as having up to thirty points. The minimum order is fifty dollars before the first-order launch discount. The listed launch discount is twenty percent, capped at fifty dollars. The work is priced by object. Two images with very different numbers of objects can require different amounts of work. For example, one thousand box objects at six cents each gives a sixty-dollar starting subtotal. That's one thousand times zero point zero six. A final quote can change with object visibility, polygon complexity, class count, image quality, and delivery requirements. The user reviews and approves the quote, then pays through the service payment workflow. Work starts after payment is confirmed. The order page shows progress using completed and total image counts. The process includes a quality-review stage before delivery. The user receives a private ZIP and has a seven-day review period. One reasonable correction round is included. Delivered files are described as available for thirty days before retention cleanup removes them. The service accepts clearly identifiable physical objects. Its policy excludes medical images and subjective atmospheric regions such as smoke, fog, rain, clouds, and shadows. So a request involving those regions falls outside the current service scope. On the owner side, there are controls for reviewing requests, preparing quotes, updating order progress, handling corrections, and delivering files. That keeps the service order organized separately from a self-service AnnoTrack project.

The last main area is the Learning Center. This connects the tools to the concepts behind them. There are four subject areas: neural networks, image processing, computer vision, and machine learning. The neural-network course runs through eighty-two lessons. It starts with the basic idea of a network and then covers loss functions, training, generalization, backpropagation, convolutional networks, segmentation, and object detection. The continuation includes topics such as transfer learning, attention, generative models, and monitoring or retraining. The other three courses each have eighty reading lessons. Each is arranged into twenty modules of four lessons. Those three courses contain two hundred and forty reading lessons in total. That's three times eighty. Adding the eighty-two neural-network lessons gives a catalog of three hundred and twenty-two lessons. That is a lesson count, not a count of completed videos. Image processing covers how an image is represented and how we change it. That includes intensity, filtering, morphology, interpolation, and image quality. Computer vision covers how we interpret images. That includes cameras, geometry, features, motion, detection, segmentation, and evaluation. Machine learning covers how we define a task, prepare data, train models, and judge their performance. A reading lesson starts with the topic and learning goals. Then it gives the explanation, a worked example, a practice task, and the reasoning behind the answer. There are extra knowledge checks and links for further reading. Narrated videos are added as production is completed. In the current registry, image processing has video entries through lesson forty. Computer vision has entries through lesson thirty-six. Machine learning also has entries through lesson thirty-six. The twelve neural-network continuation videos are registered too. A reading-course video appears on the page when its required video, poster, captions, and transcript are available. So a reading lesson can already be useful while its video is still forthcoming. The eighty reading lessons in a new course shouldn't be described as eighty completed videos. Where a video is available, the viewer can play it, turn on optional English captions, and open the matching narration transcript. Captions start off. The page also has Previous and Next links so the viewer can continue through the course. At the end of a module, there is a five-question multiple-choice quiz. Each question has four choices. After submission, the learner gets the score, correct answers, explanations, and links back to the relevant lessons. The practice target is eighty percent. With five questions, four correct answers reach that target. Four divided by five is zero point eight, or eighty percent. Guests can practice. Signed-in users can save their result to their personal learning progress. The further-reading links point to books and learning resources. The explanations and examples are presented as Tools in Vision teaching material. The references give learners places to study the subject in more depth. One possible lab activity would be to combine the Learning Center with AnnoTrack. Students could start with a short video and extract a small set of frames. They could define the classes, annotate the images, and ask another student to review the labels. Then they could inspect the dataset checks and export the result. The lessons could support the ideas used in that exercise, such as image coordinates, segmentation, validation, or the difference between training and testing. The website also has a Support Center. Users can send technical or account requests, see their requests, and reply to support. There is a place for suggestions too. The owner has a support inbox for following up with users. The account security controls show email verification, sign-in method, recognized devices, and login history. A user can sign out another recognized device. If they don't recognize a device, there is a separate secure-account flow. Permissions are checked on the server because hiding a button in the browser by itself wouldn't be enough. Underneath the website, the main application uses Python and Flask. The annotation interface uses JavaScript and a browser canvas. The database stores projects, saved labels, team membership, review records, and job status. Private storage holds image files and generated or exported outputs. Model training, media processing, generation, and ZIP export use background-job workflows. That lets the interface show progress while heavier work happens separately. OpenCV handles parts of tracking and image analysis. FFmpeg handles video operations. The project-training workflow uses YOLO. The creation tools use their configured generation or voice services. The website combines these existing libraries, models, and services with the project interface, saved annotation workflow, collaboration features, account access, and job management. When I evaluate the AI features, I would still separate a working tool from a research performance claim. A completed training job tells me that a model was produced and activated. To claim a particular accuracy or improvement, I would need a defined test set, a clear evaluation procedure, and the actual measured results. For me, the useful part of the website is being able to take a dataset through these steps in an organized way. I can prepare the files, label an object, use tracking to propose the next label, correct that proposal, ask someone to review it, and export the result for the next stage of the project. If a processing service is unavailable, that can affect a new training or generation job. Manual annotation and the saved dataset workflow can still be demonstrated independently of completing a new training run. The parking-lot example shows how the different parts fit together. We start with a video, choose useful frames, define the classes, label the cars, review the tracking suggestions, check the annotations, and export the dataset. The other tools help prepare media, create supporting content, request annotation work, or understand the methods we're using. That's the full walkthrough of Tools in Vision.