---
license: cc-by-nc-4.0
tags:
  - object-detection
  - multi-object-tracking
  - thermal
  - wildlife
  - yolo
  - rf-detr
library_name: pytorch
---

# BAMBI Wildlife MOT — trained weights

Trained model checkpoints for the **BAMBI Wildlife MOT** project: detection and
multi-object tracking of wildlife (Wild boar, Red deer, Roe deer) in aerial
**thermal** imagery from the [BAMBI dataset](https://www.bambi.eco/).

Code, documentation, and reproduction instructions:
**https://github.com/Navid-alt/BAMBI_MOT**

These weights are git-ignored in the source repo because of their size and are
hosted here instead. File paths below mirror where each script writes/expects the
checkpoint in the repo checkout.

## Files

### Detectors

| Model | File | Place at (in repo checkout) |
|---|---|---|
| YOLO26-s (640) | `yolo26-s/runs/train_yolo26-s/weights/best.pt` | `detection_models/yolo26-s/runs/train_yolo26-s/weights/best.pt` |
| YOLO26-l (1024) | `yolo26-l/runs/train_yolo26-l_1024/weights/best.pt` | `detection_models/yolo26-l/runs/train_yolo26-l_1024/weights/best.pt` |
| YOLO26-l (default res) | `yolo26-l/runs/train_yolo26-l/weights/best.pt` | `detection_models/yolo26-l/runs/train_yolo26-l/weights/best.pt` |
| RF-DETR-l — EMA (recommended) | `rf-detr-l/output/checkpoint_best_ema.pth` | `detection_models/rf-detr-l/output/checkpoint_best_ema.pth` |
| RF-DETR-l — regular | `rf-detr-l/output/checkpoint_best_regular.pth` | `detection_models/rf-detr-l/output/checkpoint_best_regular.pth` |

`*/weights/last.pt` (YOLO) and `rf-detr-l/output/{checkpoint_39,last}.ckpt`
(full training state incl. optimizer, ~527 MB each) are included for full
reproducibility but are **not needed for inference**.

### Learned tracker (TrackAssociator, MOTRv2-style)

| File | Notes |
|---|---|
| `transformer_tracking/output/associator_last.pth` | Final associator — use this for inference |
| `transformer_tracking/output/associator_epoch{0..9}.pth` | Per-epoch checkpoints |

The associator runs on top of the **frozen** RF-DETR-l detector
(`checkpoint_best_ema.pth` above).

## Classes

`0: Wild boar` · `1: Red deer` · `2: Roe deer`

## Usage

Download a file and drop it at the path listed above; the detection/tracking
scripts pick it up automatically. Example:

```python
from huggingface_hub import hf_hub_download

ckpt = hf_hub_download(
    repo_id="NavidGh/BambiMot",
    filename="rf-detr-l/output/checkpoint_best_ema.pth",
)
```

See the GitHub repo's `detection_models/readme.md` and
`transformer_tracking/README.md` for training and evaluation commands.