bowenc0221 / panoptic-deeplab

This is Pytorch re-implementation of our CVPR 2020 paper "Panoptic-DeepLab: A Simple, Strong, and Fast Baseline for Bottom-Up Panoptic Segmentation" (https://arxiv.org/abs/1911.10194)
Apache License 2.0
587 stars 117 forks source link

The poor Instance Segmentation performance on COCO #65

Closed ghost closed 3 years ago

ghost commented 3 years ago

Hi, I am curious why the Instance segmentation performance on Cityscapes seems acceptable, but that on COCO is so poor?

What do you think may be the reason for this? Thank you in advance.

bowenc0221 commented 3 years ago

Few thoughts:

  1. COCO has larger range of objects, so FPN is important. Current Panoptic-DeepLab only uses a single scale feature for detection. I would assume using FPN can improve performance a lot on COCO.
  2. AP metric itself is in favor of high recall. This is especially the case for COCO dataset. That's why top-down methods like Mask R-CNN shine on COCO.
  3. I did not find the optimal training parameters on COCO.
ghost commented 3 years ago

Few thoughts:

  1. COCO has larger range of objects, so FPN is important. Current Panoptic-DeepLab only uses a single scale feature for detection. I would assume using FPN can improve performance a lot on COCO.
  2. AP metric itself is in favor of high recall. This is especially the case for COCO dataset. That's why top-down methods like Mask R-CNN shine on COCO.
  3. I did not find the optimal training parameters on COCO.

Thank you so much!