search

pse-ecn / pose-sensitive-embedding

Pose Sensitive Embedding for Person Re-Identification (PSE)
112 stars 36 forks source link

where is the 'three consecutive convo-lutions with step sizes 2, 2, and 5' ? #36

Closed datar001 closed 4 years ago

datar001 commented 4 years ago

hi , when i read the code, i can't find the the 'three consecutive convo-lutions with step sizes 2, 2, and 5' mentioned in the paper, could you tell me where it was called。it can't be found in the resnet_v1_views.py. Thanks very much

datar001 commented 4 years ago

This the original paragraph in the paper. image

ssarfraz commented 4 years ago

Look here https://github.com/pse-ecn/pose-sensitive-embedding/blob/master/nets/resnet_utils_views.py#L199

datar001 commented 4 years ago

thanks very much. besides, i have a question Whether the number of input channels affects the model and how?. I utilize the Joint Body Parsing with 19 channels as fine-grained information, making input images with 22 channels. But the final performance of the Pose Map Models is not particularly well (map:57.9, rank1:80.7), lower than the results i get in the exposed method(map:64.1, rank1:84.1). can you give me some suggestions? or what i should watch out for?

ssarfraz commented 4 years ago

I guess utilizing more joint locations in the input should not worsen the performance.. Not sure but may be try fine tuning the conv1 first sufficiently before fine tuning the full model.

datar001 commented 4 years ago

I'm sorry to bother again. can you tell me where is the 'pose_maps/poses_maps.npy' in the RapPoseMapsDataset.py? it looks like lossing the file when i fine-tune the View Predictor in PSE model. I have generated the pose-map-file of RAP. Thanks very much.