Open aragakiyui611 opened 8 months ago
Hi, yes you're correct. The model of the original paper is trained only with IH2.6M (H) + MSCOCO. The checkpoint of this repo is trained with IH2.6M (H+M) + MSCOCO, which gives
bbox IoU: 86.25
MRRPE: 26.74 mm
MPVPE for all hand sequences: 11.55 mm MPVPE for single hand sequences: 9.58 mm MPVPE for interacting hand sequences: 12.14 mm
MPJPE for all hand sequences: 13.65 mm MPJPE for single hand sequences: 12.75 mm MPJPE for interacting hand sequences: 14.55 mm
Let me update the arxiv paper.
Is the result of HIC on the paper is trained with InterHand26M(H+M) and coco? Left is paper and right is my result (H+M). Thanks!
All experimental results of the paper are from checkpoints trained on IH2.6M (H) + MSCOCO
I reproduced the result and on HIC is worse especially MRRPE. However on the Interhand is better than on the paper. May be the random seed? or on the final epoch selecting? I test with snapshot_6.pth
Dose batch size affects? I use batch size 32 with 2 GPUs. My MRRPE on HIC is only 40.11 mm
I haven't tested with different batch size.. sorry
addict 2.4.0 certifi 2023.7.22 charset-normalizer 3.3.0 chumpy 0.70 contourpy 1.1.1 cycler 0.12.0 Cython 3.0.4 easydict 1.10 einops 0.7.0 filelock 3.12.4 fonttools 4.43.0 fsspec 2023.9.2 fvcore 0.1.5.post20221221 gitdb 4.0.11 GitPython 3.1.40 huggingface-hub 0.18.0 idna 3.4 importlib-metadata 6.8.0 importlib-resources 6.1.0 iopath 0.1.10 json-tricks 3.17.3 kiwisolver 1.4.5 kornia 0.7.0 matplotlib 3.7.3 mmcv-full 1.7.1 mmpose 0.28.0 munkres 1.1.4 numpy 1.24.4 opencv-python 4.8.1.78 packaging 23.2 pandas 2.0.3 Pillow 10.0.1 pip 23.2.1 platformdirs 3.11.0 plyfile 1.0.1 portalocker 2.8.2 psutil 5.9.6 py-cpuinfo 9.0.0 pycocotools 2.0.7 pyparsing 3.1.1 python-dateutil 2.8.2 pytorch3d 0.7.4 pytz 2023.3.post1 PyYAML 6.0.1 requests 2.31.0 safetensors 0.4.0 scipy 1.10.1 seaborn 0.13.0 setuptools 68.2.2 six 1.16.0 smmap 5.0.1 smplx 0.1.28 tabulate 0.9.0 termcolor 2.3.0 thop 0.1.1.post2209072238 timm 0.9.7 tomli 2.0.1 torch 1.12.1+cu113 torchaudio 0.12.1+cu113 torchgeometry 0.1.2 torchvision 0.13.1+cu113 tqdm 4.66.1 trimesh 4.0.0 typing_extensions 4.8.0 tzdata 2023.3 ultralytics 8.0.200 urllib3 2.0.6 wheel 0.41.2 xtcocotools 1.14.3 yacs 0.1.8 yapf 0.40.2 zipp 3.17.0
I use 2 3090 gpus with training batch size 32. So the global batch size is 32 * 2. May I know your pytorch version info and the global batch size you use?
I reproduce the most recent version model, which trained with InterHand26M (H+M) and COCO dataset. However, I found that reproduced results are better than that in the paper. Did you use human_aid only (H) InterHand26M in the paper?