open-mmlab / mmfashion

Open-source toolbox for visual fashion analysis based on PyTorch
https://open-mmlab.github.io/
Apache License 2.0
1.27k stars 283 forks source link

get_img_tensor randomScaleAndCrop #117

Open Dogacel opened 3 years ago

Dogacel commented 3 years ago

I have noted that while testing, images are being distorted. The colors are not important I know they are normalized. But the problem is for example this image is so cropped that it is almost impossible to detect it as a pants. I am not sure if the training data is trained on such samples because if it did, then you should reconsider retraining model with correct transformations. This distortion causes too much information loss IMO.

result_cropped

78 might be related to this fact.

Also only the vgg_16 global pooling model is working on attribute and category prediction. Other models are very poorly optimized. They are predicting very absurd results.