DeepLabCut
DeepLabCut provides markerless pose estimation for tracking animal body parts in videos and images. Biom supports both inference with pre-trained models and training custom models.Input formats
.mp4, .avi, .mov, .mkv, .tiff, .png, .jpg
Pre-trained models
| Model | Description |
|---|---|
| full_mouse | Full-body mouse tracking (side view) |
| full_rat | Full-body rat tracking |
| top_mouse | Top-down mouse tracking |
| side_mouse | Side-view mouse tracking |
| primate_face | Primate facial landmark tracking |
| horse | Horse body tracking |
| custom | Your own trained model (upload a DLC project or snapshot) |
Parameters
| Parameter | Range | Default | Description |
|---|---|---|---|
| Confidence cutoff (pcutoff) | 0–1 | 0.6 | Minimum confidence to show a keypoint |
| Batch size | 1–64 | 8 | Frames processed per batch |
| Num animals | 1–20 | 1 | Number of animals to track |
| Identity tracking | Toggle | Off | Track individual identity across frames |
| Overlay video | Toggle | On | Generate video with skeleton overlay |
| CSV export | Toggle | On | Export keypoint coordinates as CSV |
| Dynamic cropping | Toggle | Off | Crop around detected animals for better accuracy |
Outputs
| Output | Format | Description |
|---|---|---|
| Keypoint tracks | H5 | Full keypoint coordinates per frame |
| Coordinates | CSV | Tabular keypoint data |
| Skeleton | JSON | Skeleton connectivity definition |
| Overlay video | MP4 | Video with skeleton visualization |
| QC metrics | JSON | Quality metrics and confidence stats |
Presets
| Preset | Description |
|---|---|
| Quick Analysis | Fast inference with default settings |
| High Quality | Higher confidence threshold, all outputs enabled |
| Multi-Animal Setup | Multi-animal mode with identity tracking |
Compute requirements
| Resource | Requirement |
|---|---|
| GPU | T4 minimum (4–6+ GB VRAM) |
| Duration | ~300 seconds baseline, scales with video length |
Training custom models
You can train custom DeepLabCut models on your own labeled data:Base architectures
| Architecture | Description |
|---|---|
| ResNet-50 | Standard, good balance of speed and accuracy |
| ResNet-101 | Higher accuracy, slower |
| EfficientNet-B0 | Efficient, lower memory |
| MobileNet-V2 | Fast, mobile-optimized |
| HRNet-W32 | High-resolution, best accuracy |
| DLCRNet-MS5 | DLC-specific architecture |
| SuperAnimal-Mouse | Pre-trained on diverse mouse data |
| SuperAnimal-Primate | Pre-trained on diverse primate data |
Training parameters
| Parameter | Range | Default | Description |
|---|---|---|---|
| Max iterations | 10k–1M | 200k | Training iterations |
| Batch size | 1–32 | 8 | Training batch size |
| Augmentation | default/tensorpack/imgaug/none | default | Data augmentation strategy |
| Transfer learning | Toggle | On | Fine-tune from pre-trained weights |
Training presets
| Preset | Iterations | Architecture | Use case |
|---|---|---|---|
| Quick | 50k | MobileNet-V2 | Rapid prototyping |
| Standard | 200k | ResNet-50 | General purpose |
| High Quality | 500k | HRNet-W32 | Publication-quality results |
Training outputs
- Trained model (ZIP)
- Checkpoint files (.pb)
- Loss curves (JSON + PNG)
- Evaluation metrics (JSON)