Robust Traffic Sign Classification Models
Pretrained checkpoints for the Two-Stage Robust Traffic Sign Classification project.
The models are trained for traffic sign recognition on GTSRB and support robustness evaluation under low-light and image-corruption conditions.
Models
Stage 1 supports seven backbones:
- ConvNeXt V2
- Swin Transformer V2
- EVA-02
- EfficientNetV2
- CAFormer
- MaxViT
- CoAtNet
Repository Structure
├── model_v1.1_convnext_gtsrb.pth
├── model_v2.1_swin_gtsrb.pth
├── model_v3.1_eva02_stage1_best.pth
├── model_v4.1_effnet_gtsrb.pth
├── model_v5.1_caformer_stage1_best.pth
├── model_v6.1_maxvit_stage1_best.pth
└── model_v7.1_coatnet_stage1_best.pth
Performance
Reported result on the official GTSRB test set:
| Metric | Score |
|---|---|
| Accuracy | 99.7941% |
| Macro F1 | 99.7329% |
| Weighted F1 | 99.7944% |
Test set: 12,630 images / 43 classes
Usage
Download the required checkpoint and use it with the training/evaluation code from the GitHub repository.
git clone https://github.com/shijunqiang27/two-stage-robust-traffic-sign-classification.git
Example:
STAGE1_CKPT=./weights/stage1/model_v1.1_convnext_gtsrb.pth \
STAGE1_META_JSON=./weights/stage1/model_v1.1_convnext_stage1_meta.json \
GTSRB_ROOT=./data/GTSRB \
python evaluation/test.py
Code
Training, evaluation, ensemble inference and Stage 2 transfer-learning code:
GitHub: https://github.com/shijunqiang27/two-stage-robust-traffic-sign-classification
Dataset
Stage 1 models are trained on the German Traffic Sign Recognition Benchmark (GTSRB).
https://www.kaggle.com/datasets/meowmeowmeowmeowmeow/gtsrb-german-traffic-sign
Notes
- These checkpoints are intended for research and evaluation.
- Keep each checkpoint together with its corresponding metadata file.
- Stage 2 can reuse the Stage 1 backbone for other single-label traffic sign classification datasets.
- Dataset and upstream pretrained-model licenses remain subject to their original terms.