search

uber-research / UPSNet

UPSNet: A Unified Panoptic Segmentation Network
Other
649 stars 120 forks source link

can't run on Mapillary vistas dataset #81

Closed yangia closed 5 years ago

yangia commented 5 years ago

I tried to train the model on Mapillary vistas dataset. However, after days' effort, I still can't make it. Here is what I've done:

  1. convert the Mapillary panoptic_2018.json to panoptic_2018_stff.json via init_coco.py with some minor changes.
  2. generate semantic gt via panoptic2semantic_segmentation.py as it is done in init_coco.sh
  3. generate instances_2018_stff.json from panoptic_2018_stff.json via panoptic2detection_coco_format.py with some changes. of course --things-only is set.
  4. train upsnet with similar configuration as upsnet_resnet50_coco_4gpu.yaml.
  5. after fixing some bugs of the compatibility of annotation formats, it still cannot run. the feedback is like this upsnet some trials:
  6. I applied the above 1-5 to COCO panoptic_{train/val}2017.json. The upsnet run smoothly.
  7. Since the bug appears at proposal_mask_target.py as it is shown in the above image, I tried to filter some small instance regions (area < 25), but in vain. could anyone give some suggestions? thanks in advance.
yangia commented 5 years ago

solved. a tiny mistake. background was ignored in num_classes.

discretecoder commented 5 years ago

hi @yangia how did you use vistas ? did you convert it to coco? was it enough just to create coco json format files

Jinming-Su commented 3 years ago

@yangia Would you share the performance trained on Mapillary vistas dataset?