Questions by using pre-optimized models

[facebookresearch]

Questions by using pre-optimized models

[gengshan-y]

Hi, could you try this file? (need to save it as opts.log in the same directory as pre-trained models). I realized the one provided in the pre-trained model link is for quadruped animals.

This runs without problem for me.

[starcxl]

Many thanks for your reply! I try the opts-human.log you provided, but the result is the same as the above.

Ah, I know what causes the difference, it's due to missing the 'init-cam' folder. If 'init-cam' is absence, the inference results are like above.

So could you please provide the 'init-cam' files matching the pre-trained models?

[gengshan-y]

It should work without the init-cam folder. That not being used at inference time.

Was there a warning when executing the code with the following message?
!!!deleting video specific dicts due to size mismatch!!!

It would be best if you attach the a log from the console. Here's mine:
log.txt

Were you able to run the cat example without a problem?

[starcxl]

Many thanks for your reply! Yes, there is a warning as you mentioned. "!!!deleting video specific dicts due to size mismatch!!!" And I got the same warning and the same strange results when running the cat example. I attach my log here: log.txt

I check the warning code here. For my situation, the states['near_far'].shape[0] is 583, while the self.model.near_far.shape[0] is 582.

I don't know what causes this mismatch. Could you please tell me the potential reasons?

[gengshan-y]

This suggests you are missing an image in database/DAVIS/JPEGImages/Full-Resolution/human-cap006/, and below is my folder (adult-500 is the old name of human-cap006)

[starcxl]

Oh! It works now if I replace human-cap006 datasets with the corresponding datasets you provided.

For the absent image issue, I retried preprocessed command below, but still, it extracts 00000-00037.jpg from human-cap_6.MOV, with 00038.jpg missing. (using ffmeng 3.4.8, pytorch 1.7.1, cuda 11.0)
bash preprocess/preprocess.sh human-cap_6 .MOV y 10

I am confused about what causes this difference. Could you please retry the command above to see if 00038.jpg could be extracted normally?

[gengshan-y]

Can confirm it works for me.

In the preprocessing script, we first use ffmpeg to extract frames, and then the frames are moved to tmp/$seqname/images.

Then preprocess/mask.py reads images and runs segmentation. The frames with good detections are saved to database/DAVIS/JPEGImages/...

Attaching the log of running mask.py below:
log.txt

Can you confirm ffmpeg or mask.py are producing expected outputs?