search

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
590 stars 117 forks source link

Cityscapes has around 30 classes, but it handles 19 #120

Closed ankur-chr closed 2 years ago

ankur-chr commented 2 years ago

Does the old code (not detectron2) work with the latest Cityscapes dataset which has around 30 classes? I noticed that the code has around 19 classes. And my results are weird with google colab, resnet34, ims_per_batch 1, 10k max iterations. The predicted images have similar coloured (reddish color tones) segmentation which is confusing me, as it seems unexpected. Is this because this is not yet post-processed yet?

Here is the prediction example from the output/...../panoptic/predictions folder (after 10k iterations) image

When I increase the max iterations to 90k, it takes a huge amount of time to train. So, midway through I checked the debug output images and compared them with the target images. It shows good results, but they are still not accurate after about around 46k iterations (still running for 90k: currently around 12 hours in the training).

Is IMS_PER_BATCH = 1 a valid choice for deep learning training considering it refers to a batch size of only 1 image?

@bowenc0221 could you please share some insights?

UPDATE: It seems with more iterations, the accuracy is improving (when comparing the training target and output images under the debug_train folder)

ankur-chr commented 2 years ago

Found the answer: 19 classes are used for evaluation purpose, even though the overall set is around 30.

https://github.com/mcordts/cityscapesScripts/issues/8

ankur-chr commented 2 years ago

Closing this thread.