Bugs record

[ruiyan1995]

Thanks for your wonderful work! I record some minor bugs as follows,

1. Could you provide the script for extracting frames? (I just used the script provided in here with 12 FPS)
2. line 184, code/train.py: no attribute 'arch', and no such directory './ckpt/xxxxx' (need to create the path of args.ckpt and delete 'arch');
3. line 284, code/data_utils/data_loader_frames.py: unnecessary?
4. line 281, code/data_utils/data_loader_frames.py: the 'global_box_id' will be larger than 'num_boxes', thus needs "try .. except ..."?
5. line 298, code/model/model_lib.py: the 'train()' function of VideoModelCoordLatentNL that causes to bad results need to be deleted.

More minor bugs will be updated here.

Some important bugs:

1. To aovid random test results, should modify return ret to return [ret[-1]] at here
2. drop_last = True at here is wrong!

[joaanna]

Hi Rui,

thank you for the bugs report! Feel free to send a pull request.

[xxxzhi]

So many problems...

[turtleman99]

For 2, and 3, I think is 'yes'.
For 4, based on the code, I believe they only use one box for one standard_category as well as hands. Are there such complicated frames?

[ruiyan1995]

@turtleman99 I found that there are more than 4 objects in some frames, but the code did not process this case.

[turtleman99]

@turtleman99 I found that there are more than 4 objects in some frames, but the code did not process this case.

Yes, I found too. To be more clear, the code only addresses only one object for each category and the maximun category number is 4 by default. My solution is to use the annotations with high confidence when collecting them.