Discrepancy in performance compared to paper

[netw0rkf10w]

Hi,

First of all, thanks for the code! I was looking for a PyTorch port of DeepLabv3+ that can use the official TensorFlow checkpoints. Your repo is the only one I could find! Great work!

There seems however to be some discrepancy compared to the TensorFlow version.
I tried the xception65_coco_voc_trainval checkpoint on the Pascal VOC test set with flipping and multi-scale ([0.5:0.25:1.75]) inference, and obtained mIoU = 85.93630 (by uploading to the evaluation server), which far inferior to the reported 87.80%.

In addition to the difference in BatchNorm's eps that you discovered previously, there should be some other details that affect the performance. Unfortunately I was unable to find them. I hope you or somebody can.

[JoOkuma]

@netw0rkf10w were you able to improve the performance to get closer to the TF version?