The output of DefaultPredictor is never normalized. However, cosine similarity calculation requires normalized embedding. This bug also causes wrong ranking/visualization as shown in a previous issue https://github.com/JDAI-CV/fast-reid/issues/651#issue-1194006183
Expected behavior:
either normalization should be done outside the predictor every time, or there should be normalization inside predictor
Instructions To Reproduce the 🐛 Bug:
https://github.com/JDAI-CV/fast-reid/blob/817c748e8c44ed8c12c3e95de355610c094fade1/fastreid/engine/defaults.py#L153
The output of DefaultPredictor is never normalized. However, cosine similarity calculation requires normalized embedding. This bug also causes wrong ranking/visualization as shown in a previous issue https://github.com/JDAI-CV/fast-reid/issues/651#issue-1194006183
Expected behavior:
either normalization should be done outside the predictor every time, or there should be normalization inside predictor
Environment:
problem exists in master branch