make the image & annotations follow below format
--OpenLaneV2
|-- av2_map_anns_val.json # used for val record the gt annotations
|-- test
|-- test_annotations.json
|-- train
|-- train_annotations.json
|-- val
`-- val_annotations.json
cd CVPR_HDMAP;
zsh tools/dist_train.sh projects/configs/custom/maptr_tiny_r50_24e_800x1024 ngpus
cd CVPR_HDMAP;
export PYTHONPATH=CVPR_HDMAP
python tools/maptr/vis_pred_av2.py config_path ckpt_path
- Image proces: Src image resize to (1550x2048) ---> scale(0.5) ----> 775x1024 ---> padding ---> 800x1024
ann_file_s3
may be comments, if you read the dataset form local
Config | Ped_crossing | Divider | Boundary | mAP | |
---|---|---|---|---|---|
maptr_tiny_r50_24e_800x1024 | 50.79 | 61.94 | 60.20 | 57.64 |