search

hkchengrex / Tracking-Anything-with-DEVA

[ICCV 2023] Tracking Anything with Decoupled Video Segmentation
https://hkchengrex.com/Tracking-Anything-with-DEVA/
Other
1.21k stars 128 forks source link

How do I reproduce the results of the rvos report？ #73

Open xinbaiw opened 5 months ago

xinbaiw commented 5 months ago

The inference is that the model being called is DEVA-propagation.pth? I used this model to infer the J&F on the rvos task, which is 64.8. The one reported in the paper is 66.0. Can you help me find out where I went wrong?

image

xinbaiw commented 5 months ago

The inference is that the model being called is DEVA-propagation.pth? I used this model to infer the J&F on the rvos task, which is 64.8. The one reported in the paper is 66.0. Can you help me find out where I went wrong?

image

image

hkchengrex commented 5 months ago

Can you compare your output with our pre-computed one and see if there is any discrepancy? It should reproduce exactly. There is too little information here for me to look into it.

xinbaiw commented 5 months ago

Can you compare your output with our pre-computed one and see if there is any discrepancy? It should reproduce exactly. There is too little information here for me to look into it.

Can you find anything wrong with these details？

parser.add_argument('--mask_path', default='/home/data2/Ref-YouTubeVOS-ReferFormer-SwinL-detections') parser.add_argument('--model', default='./saves/DEVA-propagation.pth') parser.add_argument('--img_path', default='/home/data2/rvos/valid/JPEGImages')

image image

xinbaiw commented 5 months ago

Hi, can you help me?

hkchengrex commented 5 months ago

I cannot peek into your file system and know what is in each folder. One potential reason is that you are not using all the frames in YouTubeVOS (i.e., https://drive.google.com/drive/folders/16uw2j3o8Pi2xhvgGopYMKyjKtzV9RQ6c)

xinbaiw commented 5 months ago

I cannot peek into your file system and know what is in each folder. One potential reason is that you are not using all the frames in YouTubeVOS (i.e., https://drive.google.com/drive/folders/16uw2j3o8Pi2xhvgGopYMKyjKtzV9RQ6c)

Thank you for your reply. When I used all frames of youtube-vos to test, out of memory, single card 4090 (24g), do you have a solution?

xinbaiw commented 5 months ago

How to reduce memory usage when testing?

hkchengrex commented 5 months ago

See here https://github.com/hkchengrex/Tracking-Anything-with-DEVA/blob/main/docs/EVALUATION.md#general-arguments

Use --amp

longmalongma commented 5 months ago

See here https://github.com/hkchengrex/Tracking-Anything-with-DEVA/blob/main/docs/EVALUATION.md#general-arguments

Use --amp

After using amp, the memory is still not enough. Is there any other way?

longmalongma commented 5 months ago

See here https://github.com/hkchengrex/Tracking-Anything-with-DEVA/blob/main/docs/EVALUATION.md#general-arguments Use --amp

After using amp, the memory is still not enough. Is there any other way?

Why does the test go out or memory at a later stage?