Pretrained Models¶
DREEM provides pretrained models for multi-object tracking across two domains: animals and microscopy. Both models are available on Hugging Face and can be used directly with the dreem CLI.
Summary¶
| Model | Domain | Training Data | Hugging Face |
|---|---|---|---|
dreem-animals-pretrained |
Animals | ~1M frames across multiple species | talmolab/dreem-animals-pretrained |
dreem-microscopy-pretrained |
Microscopy | ~100K frames across cells and organelles | talmolab/dreem-microscopy-pretrained |
Animals¶
A general-purpose animal identity tracking model trained on ~1 million frames of proofread, identity-tracked public data spanning multiple species and scales, from fruit flies to mice. Note that you do not need the config to train the model; the CLI has many options that work without having to use a config file. See CLI documentation for available options.
| Input format | Videos with detection labels in .slp format |
| Training data | ~1M frames (datasets) |
| Hardware | 4x A40 GPUs |
| Metrics | CLEARMOT (py-motmetrics) |
| Training config | animals-pretrained-config.yaml |
| Download | animals-pretrained.ckpt |
Microscopy¶
A general-purpose microscopy identity tracking model trained on ~100K frames of proofread, identity-tracked public data spanning diverse biological scales, from organelles to cell nuclei.
| Input format | Videos with detection labels in .slp or Cell Tracking Challenge format |
| Training data | ~100K frames (datasets) |
| Hardware | 4x A40 GPUs |
| Metrics | CLEARMOT (py-motmetrics), CTC (py-ctcmetrics) |
| Config | microscopy-pretrained-config.yaml |
| Download | pretrained-microscopy.ckpt |