TCSR / README.md
wodanile's picture
Update TCSR model card
e1d5b2b verified
|
Raw History Blame Contribute Delete
3.99 kB
---
license: other
library_name: pytorch
tags:
- object-detection
- aerial-object-detection
- small-object-detection
- visdrone
- rtdetrv2
- dfine
- tcsr
---
# TCSR VisDrone Checkpoints
This repository contains the public VisDrone checkpoints for **Shallow-Response Calibration for Real-Time Aerial Small Object Detection**. The implementation, configurations, training commands, and evaluation tools are maintained in the [TCSR GitHub repository](https://github.com/txu8847-wq/TCSR).
Because of storage and long-term maintenance constraints, the hosted checkpoint collection is limited to VisDrone. The AI-TOD-v2 experiment configurations remain available in the GitHub repository.
## Available Checkpoints
| Framework | Backbone | Variant | VisDrone AP | AP gain | File |
|:--|:--|:--|--:|--:|:--|
| D-FINE | HGNetv2-S | Baseline | 27.56 | - | `dfine_hgnetv2_s_visdrone_baseline_ap=27.56.pth` |
| D-FINE | HGNetv2-S | Direct | 28.51 | +0.95 | `dfine_hgnetv2_s_visdrone_detection_direct_hard_ap=28.51.pth` |
| D-FINE | HGNetv2-S | TCSR-Light | 29.11 | +1.55 | `dfine_hgnetv2_s_visdrone_TCSR_light=64_ap=29.11.pth` |
| D-FINE | HGNetv2-S | TCSR-Hard | 30.71 | +3.15 | `dfine_hgnetv2_s_visdrone_detection_TCSR_hard_ap=30.71.pth` |
| RT-DETRv2 | HGNetv2-X | Baseline | 26.52 | - | `rtdetrv2_hgnetv2_x_visdrone_baseline_72e_seed0_ap=26.52.pth` |
| RT-DETRv2 | HGNetv2-X | SE-Direct | 28.07 | +1.55 | `rtdetrv2_hgnetv2_x_visdrone_se_direct_72e_seed0_ap=28.07.pth` |
| RT-DETRv2 | HGNetv2-X | TCSR-Hard (deterministic) | 29.37 | +2.85 | `rtdetrv2_hgnetv2_x_visdrone_tcsr_det_72e_seed=0_ap=29.37.pth` |
| RT-DETRv2 | HGNetv2-X | TCSR-Hard | 29.30 | +2.78 | `rtdetrv2_hgnetv2_x_visdrone_TCSR_72e_seed0_AP=29.30.pth` |
| RT-DETRv2 | ResNet-18 | Baseline | 29.40 | - | `rtdetrv2_r18vd_visdrone_baseline_200e_seed0_ap=29.40.pth` |
| RT-DETRv2 | ResNet-18 | Direct | 29.76 | +0.36 | `rtdetrv2_r18vd_visdrone_direct_200e_seed0_ap=29.76.pth` |
| RT-DETRv2 | ResNet-18 | TCSR-Hard | 31.17 | +1.77 | `rtdetrv2_r18vd_200e_visdrone_TCSR_hard=256_ap=31.17.pth` |
The table lists the checkpoint files currently hosted in this repository. The revised paper reports the complete RT-DETRv2/ResNet-18 sequence: Baseline 29.40 AP, Direct 29.76 AP (+0.36), TCSR-Light 30.13 AP (+0.73), and TCSR-Hard 31.17 AP (+1.77). Results for additional variants are summarized in the [TCSR GitHub repository](https://github.com/txu8847-wq/TCSR).
For RT-DETRv2/HGNetv2-X, SE-Direct is the matched global channel-recalibration control. It uses standard squeeze-and-excitation after `C2` alignment, a reduction ratio of 16, no stochastic response exposure, the 72-epoch schedule, automatic mixed precision, and seed `0`. Its matching configuration is `RT-DETR/rtdetrv2_pytorch/configs/rtdetrv2/rtdetrv2_hgnetv2_x_visdrone_se_direct.yml` in the GitHub repository.
The deterministic TCSR-Hard checkpoint retains the complete 384-channel local TCSR transformation but disables stochastic response exposure (`tcsr_exposure_std: 0.0`). It uses the same 72-epoch HGNetv2-X setup, automatic mixed precision, and seed `0`; its matching GitHub configuration is `RT-DETR/rtdetrv2_pytorch/configs/rtdetrv2/rtdetrv2_hgnetv2_x_visdrone_tcsr_det.yml`.
## Usage
Download one checkpoint with the Hugging Face CLI:
```bash
hf download wodanile/TCSR <checkpoint-name>.pth --local-dir ./checkpoints
```
Refer to the GitHub repository for the matching configuration and use the framework's evaluation or resume option when loading a trained VisDrone checkpoint. The RT-DETRv2 and D-FINE loaders include compatibility remapping for checkpoints produced before the public TCSR namespace cleanup.
## Important Notes
- These are PyTorch checkpoint files. Load checkpoints only from sources you trust.
- Checkpoint filenames record the framework, backbone, variant, schedule or seed where applicable, and reported validation AP.
- The COCO initialization checkpoint `rtdetrv2_r18vd_120e_coco_rerun_48.1.pth` is not part of this repository.