Huge difference between N.A., flipping and multiscale testing

[pikkuve]

I've been training on a relatively small dataset starting from COCO-pretrained weights (dataset specs: about 5k images, over 3 classes with approx. 4000, 3000 and 24000 bounding boxes per object class). I focused on object detection, and used a lr =0.0003125 for 140 epochs with reduction at steps 80, 100, 125 and the results for N.A. testing, flipping and multiscale testing are wildly different, with AP50 at about 0.30-0.40 for N.A. and 0.60-70 for flipping and around 0.8-0.9 for multiscale testing. Any idea where this could stem from? Similar setting on a comparable dataset has different behavior: with an improvement from N.A. to F. and M.S. but nothing too drastic.