Training results in detections at all anchorboxes

[aaronrmm]

After training on the VOC2007 dataset with default options except epochs=5, and efficientdet-d2, I am getting detections of all classes at all anchorboxes when using both eval.py and demo.py. Changing thresholds to 0.99+ does not curb the detections.

I get the same result when training with a very tiny dataset with fewer epochs. Additionally my validation mAP is always 0, even when I evaluate on the training set.

I am using Luke Melas's backbone as suggested in this issue#111.

• Relevant packages in my conda environment are:

• Name Version Build Channel

• albumentations == 0.4.5 == pypi_0 pypi
• cudatoolkit == 10.1.243 == h6bb024c_0
• numpy == 1.18.1 == py38h4f9e942_0
• numpy-base == 1.18.1 == py38hde5b4d6_1
• opencv-contrib-python == 4.2.0.32 == pypi_0 pypi
• pytoan == 0.6.4 == pypi_0 pypi
• pytorch == 1.4.0 == py3.8_cuda10.1.243_cudnn7.6.3_0 pytorch
• torchvision == 0.5.0 == py38_cu101 pytorch

Here is the output for python demo.py --cam -w=./saved/weights/VOC/efficientdet-d2/checkpoint_5.pth -t=.999 -it=.999 :