---
license: mit
library_name: pytorch
pipeline_tag: object-detection
tags:
  - object-detection
  - instance-segmentation
  - keypoint-detection
  - semantic-segmentation
  - depth-estimation
  - oriented-object-detection
  - real-time
  - linear-attention
---

# GTR: Gated Token Recurrence for Efficient Dense Prediction

Checkpoints for [GTR](https://github.com/Intellindust-AI-Lab/GTR). Code, configs, evaluation and
training instructions are in the GitHub repository.

```bash
git clone https://github.com/Intellindust-AI-Lab/GTR.git && cd GTR
hf download Phoenix8125/GTR --local-dir weights
python train.py -c configs/det/coco_finetune/gtr_s.yml --test-only -r weights/det/gtr_s_coco.pth
```

Each task checkpoint is a dict `{'model': raw weights, 'ema': {'module': EMA weights, ...}, 'last_epoch': int}`.
All numbers below are measured with the EMA weights.

| Folder | Task / benchmark | Metric | S | M | L | X |
|---|---|---|---|---|---|---|
| `det/` | detection, COCO val2017 | AP | 53.6 | 57.3 | 58.9 | 59.4 |
| `seg/` | instance segmentation, COCO val2017 | mask AP | 45.0 | 47.7 | 49.5 | 49.8 |
| `pose/` | human pose, COCO val2017 | AP | 70.1 | 74.1 | 74.7 | 75.5 |
| `semseg/` | semantic segmentation, Cityscapes val | mIoU | 81.5 | 83.0 | 83.2 | 83.6 |
| `depth/` | monocular depth, NYU Depth V2 (Eigen test) | δ1 / AbsRel | 0.946 / 0.074 | 0.952 / 0.069 | 0.951 / 0.069 | 0.954 / 0.067 |
| `obb/` | oriented detection, DOTA-v1.0 test | AP50 | 80.0 | – | – | 81.3 |
| `obj365/` | Objects365 pre-trained detectors | – | ✓ | ✓ | ✓ | ✓ |