supervisely · Oleg-1703 · Oct 9, 2025 · Oct 9, 2025 · Oct 23, 2025 · Nov 3, 2025
diff --git a/.gitbook/assets/auto-track.jpg b/.gitbook/assets/auto-track.jpg
diff --git a/.gitbook/assets/neural-networks/predict_app_from_datasets.jpg b/.gitbook/assets/neural-networks/predict_app_from_datasets.jpg
diff --git a/.gitbook/assets/neural-networks/predict_app_from_projects.jpg b/.gitbook/assets/neural-networks/predict_app_from_projects.jpg
diff --git a/.gitbook/assets/neural-networks/predict_app_tracking_settings.jpg b/.gitbook/assets/neural-networks/predict_app_tracking_settings.jpg
diff --git a/.gitbook/assets/neural-networks/tracking_labeling_tool.jpg b/.gitbook/assets/neural-networks/tracking_labeling_tool.jpg
diff --git a/SUMMARY.md b/SUMMARY.md
@@ -130,6 +130,7 @@
   * [Labeling Quality Control](labeling/jobs/Labeling-Quality-Control.md)
 * [Labeling Performance](labeling/labeling-performance.md)
 * [Labeling with AI-Assistance](labeling/AI-labeling/AI-Labeling.md)
+* [Automatic Video Object Tracking](labeling/AI-labeling/auto-tracking.md)
 
 ## 🤝 Collaboration
 

diff --git a/labeling/AI-labeling/auto-tracking.md b/labeling/AI-labeling/auto-tracking.md
@@ -0,0 +1,79 @@
+# Automatic Video Object Tracking
+
+Object tracking is a technology that automatically monitors object positions in each video frame. Instead of manually labeling thousands of frames, you can use AI models to automate this process. Auto-tracking reduces labeling time by tens of times, ensuring consistent quality and allowing you to process large volumes of data without proportionally increasing your annotation team.
+
+Supervisely offers two main approaches to auto-tracking, each suited for different scenarios.
+
+---
+
+## Two Approaches to Video Object Tracking
+
+There are two fundamentally different ways to track objects in videos, and they differ in how you start the tracking process.
+
+**Tracking-by-Detection** works on completely unlabeled videos. A detector (like YOLO or RT-DETRv2) runs on every frame to find objects of predefined classes, then a tracking algorithm links these detections across frames and assigns each object a unique ID. You press one button on an empty video and get automatic annotations for all objects of the classes your detector can recognize.
+
+**First Frame Initialized Tracking** requires you to manually annotate objects first. You create annotations on the objects you want to track, and then the tracking model follows these specific objects through subsequent frames. The model uses your initial annotations as templates and tracks exactly those objects you selected, regardless of their class.
+
+The key difference: tracking-by-detection creates annotations from scratch automatically on unlabeled video, while first frame initialized tracking continues tracking from your manually created annotations.
+
+---
+
+## 1. Tracking-by-Detection
+
+### What It Is
+
+Tracking-by-Detection is an approach where a detector runs on each video frame to detect objects of predefined classes, and then a tracking algorithm (e.g., BoT-SORT) links these detections between frames and assigns each object a unique ID. The detector finds objects in each frame, the tracker links detections between frames and forms object trajectories, assigning each a unique ID that persists throughout the video.
+
+### When to Use
+
+This method is suitable when you have a detection model (custom-trained or pre-trained) that can predict the object classes you need to track. It works well for scenes with multiple objects of known classes, automatically handles new objects appearing in the frame, and deals with occlusions and scale changes. Best for tasks where you want to automatically label all objects of specific classes without manually annotating each one.
+
+### How to Use
+
+The most convenient way for annotators is to launch tracking-by-detection directly from the labeling tool:
+
+1. Deploy an object detection or segmentation model ([how to deploy a model](../../neural-networks/inference-and-deployment/supervisely-serving-apps.md#how-to-deploy-a-model))
+2. Open the video in the labeling tool and click the blue Track button at the top of the interface
+3. In the "tracking-by-detection engine" field, select your deployed model and click the "tracking-by-detection" button
+4. Annotations with tracks will automatically load into your video
+
+![Tracking in labeling tool](../../.gitbook/assets/neural-networks/tracking_labeling_tool.jpg)
+
+Besides the labeling tool, tracking-by-detection is available through [Predict APP](https://ecosystem.supervisely.com/apps/apply-nn) — a universal application for applying neural networks to video, through [API](../../neural-networks/inference-and-deployment/video-object-tracking.md#option-4-tracking-via-api) for integration into your pipeline, or for [local execution](../../neural-networks/inference-and-deployment/video-object-tracking.md#option-5-run-tracker-locally) in your code via Supervisely SDK.
+
+For a complete guide on tracking-by-detection, parameter configuration, and usage, see the [Video Object Tracking](../../neural-networks/inference-and-deployment/video-object-tracking.md) documentation. Additional information about working in the labeling tool can be found in the [Video Labeling Tool documentation](../labeling-toolbox/videos-3.0.md).
+
+---
+
+## 2. First Frame Initialized Tracking
+
+### What It Is
+
+First Frame Initialized Tracking in Supervisely is implemented through the [Auto Track app](https://ecosystem.supervisely.com/apps/auto-track), which uses class-agnostic tracking models. These models can track any object based on your initial annotation.
+
+The workflow is straightforward: you create annotations (bounding boxes, masks, or skeletons) on the objects you want to track on the current frame, and the model automatically tracks these specific objects through subsequent frames. The neural network uses your annotations as templates and follows the visual characteristics of the selected objects. Class-agnostic means the model doesn't need to know what you're tracking (person, car, animal) — it simply follows the visual features of whatever you've annotated.
+
+### When to Use
+
+This approach is ideal when you need to track specific individual objects rather than all objects of a class. For example, tracking "this particular red car" rather than "all cars", tracking objects without a ready-made detection model, or tracking a small number of specific objects (1-20) in the frame. Also useful for correcting tracking errors — when you need to fix a mistake and re-track an object from the current frame forward.
+
+### How to Use
+
+1. Open [Auto Track app](https://ecosystem.supervisely.com/apps/auto-track) from the Supervisely Ecosystem and click "Run Application" button
+2. Select Neural Network models for the geometries you want to track
+   - You can also deploy a new model directly from the app UI by clicking "Deploy new model" button, selecting the model, agent, and clicking "Deploy model"
+3. Start tracking in Video Labeling Tool:
+   - Open your video in Video Labeling Tool
+   - Create annotations on the objects you want to track
+   - Select "Auto-Track" application in the tracking engine dropdown at the top of the interface
+   - The app will automatically track objects when you create or edit them, or click the "track" button to start tracking all objects in the current frame
+
+![Auto-track in labeling tool](../../.gitbook/assets/auto-track.jpg)
+
+For detailed information about Auto Track features and settings, see the [Auto Track app documentation](https://ecosystem.supervisely.com/apps/auto-track). Additional information about the annotation workflow is available in the [Video Labeling Tool](../labeling-toolbox/videos-3.0.md) documentation.
+
+---
+
+## Additional Resources
+
+Complete documentation on [Video Object Tracking](../../neural-networks/inference-and-deployment/video-object-tracking.md) contains detailed descriptions of tracking-by-detection, tracker parameters, and API. A guide to working in the modern video labeling interface can be found in the [Video Labeling Tool 3.0](../labeling-toolbox/videos-3.0.md) documentation. An overview of tracking types and available models is in [Video Tracking Overview](../videos/video-tracking.md). A catalog of all available models and applications is in [Supervisely Ecosystem](https://ecosystem.supervisely.com/).
diff --git a/neural-networks/inference-and-deployment/video-object-tracking.md b/neural-networks/inference-and-deployment/video-object-tracking.md
@@ -37,17 +37,42 @@ Additionally, we found some improvements in the original codebase and fixed mino
 
 ## How to Use Tracker
 
-Users have three ways to apply a tracker in different scenarios:
+Users have four ways to apply a tracker in different scenarios:
 
-1. **[Apply NN to Video](#option-1-apply-nn-to-video)**: the most convenient way is to use our [Apply NN to Video](https://ecosystem.supervisely.com/apps/apply-nn-to-videos-project) app on the Supervisely platform.
-2. **[Tracking via API](#option-2-tracking-via-api)**: programmatic access through the python API - you send requests and receive predictions while tracking runs on the server.
-3. **[Run Tracker Locally](#option-3-run-tracker-locally)**: use the tracker inside your own code or application with the Supervisely SDK on your machine.
+1. **[Predict APP](https://ecosystem.supervisely.com/apps/apply-nn)**: the most convenient way is to use our [Predict APP](https://ecosystem.supervisely.com/apps/apply-nn) app on the Supervisely platform.
+2. **[Apply NN to Video (Legacy)](#option-1-apply-nn-to-video)**: an additional option to apply neural networks to videos is to use our [Apply NN to Video](https://ecosystem.supervisely.com/apps/apply-nn-to-videos-project) app on the Supervisely platform.
+3. **[From labeling tool](../../labeling/labeling-toolbox/videos-3.0.md)**: run object tracking directly from the labeling tool to automatically pre-annotate your videos.
+4. **[Tracking via API](#option-2-tracking-via-api)**: programmatic access through the python API - you send requests and receive predictions while tracking runs on the server.
+5. **[Run Tracker Locally](#option-3-run-tracker-locally)**: use the tracker inside your own code or application with the Supervisely SDK on your machine.
 
 In this guide we will cover all three options.
 
 ---
+### Option 1: Predict APP 
 
-### Option 1: Apply NN to Video
+Launch [Predict APP](https://ecosystem.supervisely.com/apps/apply-nn) app in the Supervisely platform, deploy a detection model, configure inference and tracking settings, and apply the model to your video project.
+
+- [Deploy a model](supervisely-serving-apps.md#how-to-deploy-a-model) for Object Detection or Instance Segmentation.  
+- Launch the **[Predict APP](https://ecosystem.supervisely.com/apps/apply-nn)** in one of the following ways:  
+   - From the **APP Ecosystem** or **Neural Networks** section on the platform.  
+   - From the **Projects** section, applied to a specific project.  
+     ![Predict App from Projects](../../.gitbook/assets/neural-networks/predict_app_from_projects.jpg)
+   - From the **Datasets** section, applied to one or multiple videos.  
+     ![Predict App from Datasets](../../.gitbook/assets/neural-networks/predict_app_from_datasets.jpg)
+- Select a project, dataset, and video, choose a model, select classes to run inference on, and enable the **Tracking** option.  
+   For more details about configuring the Predict APP, see the [documentation](https://ecosystem.supervisely.com/apps/apply-nn).  
+- In **Advanced Settings**, you can specify hyperparameters (see [Hyperparameter Configuration](#hyperparameter-configuration)).  
+- Click **RUN** to start the inference process.
+
+<figure><img src="../../.gitbook/assets/neural-networks/predict_app_tracking_settings.jpg" alt="Apply NN with tracker screenshot"><figcaption>Tracking settings</figcaption></figure>
+
+
+### Option 2: Apply NN to Video *(Legacy)*
+
+> ⚠️ **Note:** This is a legacy option kept for backward compatibility.  
+> We now recommend using the **[Predict APP](#option-1-predict-app)** instead.  
+>  
+> You can still use the **[Apply NN to Video](https://ecosystem.supervisely.com/apps/apply-nn-to-videos-project)** app in Supervisely.
 
 Launch [Apply NN to Video](https://ecosystem.supervisely.com/apps/apply-nn-to-videos-project) app in the Supervisely platform, deploy a detection model, configure inference and tracking settings, and apply the model to your video project.
 
@@ -75,7 +100,24 @@ track_buffer: 30
 
 > Read more about applying neural networks in [**Supervisely Serving Apps**](supervisely-serving-apps.md).
 
-### Option 2: Tracking via API
+
+### Option 3: Tracking via Labeling tool
+
+This method allows you to start tracking-by-detection right inside the labeling tool. It helps you automatically create preliminary annotations of objects and their tracks in a video before you begin manual refinement.
+
+1. [Deploy a model](supervisely-serving-apps.md#how-to-deploy-a-model) for Object Detection or Instance Segmentation.
+2. Open the [labeling tool](../../labeling/labeling-toolbox/videos-3.0.md) from your video dataset on the platform.
+3. Open the Track section by clicking the blue button at the top of the interface.
+4. In the tracking-by-detection engine field, select the model you deployed in step 1.
+5. Click the “tracking-by-detection” button to start the tracking process.
+6. The annotations with tracks will be automatically loaded into your video opened in the labeling tool.
+
+<figure><img src="../../.gitbook/assets/neural-networks/tracking_labeling_tool.jpg" alt="tracking-by-detection in labeling tool"><figcaption>Tracking via labeling tool</figcaption></figure>
+
+
+> Read more about labeling tool in [**Documentation**](../../labeling/labeling-toolbox/videos-3.0.md).
+
+### Option 4: Tracking via API
 
 > This section is based on the [Prediction API](prediction-api.md), check it for more details.
 
@@ -136,7 +178,7 @@ output_path = "output.mp4"
 visualize(predictions, video_path, output_path)
 ```
 
-### Option 3: Run Tracker Locally
+### Option 5: Run Tracker Locally
 
 This approach allows you to use the tracker inside your own code or application using the Supervisely SDK on the same machine/hardware, without sending requests to the server.