Very Low mAP on COCO 2014 When Training from Scratch Using darknet53.conv.74 Pretrained Weights

Below is the code version I used. I followed the steps in the README, including testing and training on the COCO 2014 dataset.

• The test mAP is only 53.913%, which is lower than the 55.5% mentioned in the README.

• The training mAP is even worse. It seems that the model fails to train properly, as the final mAP is only about 10% after completing 300 epochs.

• When I modify the initial learning rate in config/yolov3.cfg to 0.00001, the mAP gradually increases at the beginning, reaching around 26%. However, after approximately 60 epochs, the mAP increases very slowly, and by epoch 300, it remains around 26%, almost the same as at epoch 60. This suggests that the learning rate is too small after 60 epochs.

• If I comment out the code that decreases the learning rate, it does not make much difference—the final mAP still remains around 26% after 300 epochs. (This is after modifying the initial learning rate in config/yolov3.cfg to 0.00001.)