Is there any problem to predict image rotation with image classification?

[lvbohui]

Problem

I used the mobilenet-v3 network to train a model to solve a four-category problem, but there is only one category in the prediction result of the final model, correspondingly, the accuracy rate is about 25%. The number of four categories in my data is not much different, about 25% in each category

Related info

1. pip list | grep "mmcv\|mmcls\|^torch" 命令的输出

mmcls                             0.22.1    /home/xiaojiu/gitlab-pro/mmclassification
mmcv-full                         1.5.0
torch                             1.10.2
torchvision                       0.11.3

2. Related files What I want to distinguish is the four directions of sink, in other words, my labels are up, down, left, right, so I removed the random flips in train_pipeline and test_pipeline. Is there a problem in predicting the rotation direction of the same object using image classification? config_sink200.zip My evaluation metric is: evaluation = dict(interval=50, metric='accuracy', metric_options={'topk': 1}) This can be confirmed at line 39 of dataset/sink200_bs32_pil_resize.py in the above file

And this is the result:result.txt

[mm-assistant[bot]]

Please use English or English & Chinese for issues so that we could have broader discussion.

[Ezra-Yu]

First, check the annotations of the dataset. you can use this tool.

[tonysy]

This issue will be closed as it is inactive, feel free to re-open it if necessary.