search

HRNet / HRNet-Semantic-Segmentation

The OCR approach is rephrased as Segmentation Transformer: https://arxiv.org/abs/1909.11065. This is an official implementation of semantic segmentation for HRNet. https://arxiv.org/abs/1908.07919
Other
3.12k stars 685 forks source link

Is "HRNetV2-W18-Small-v2" SOTA for lightweight segmentation? #133

Closed InternetMaster1 closed 4 years ago

InternetMaster1 commented 4 years ago

Wow congratulations for the amazing library!

I am looking to implement high-quality semantic segmentation on a mobile device for accurate human segmentation for still images (i.e. non-realtime).

1) SOTA for light-weight?

I have been researching a lot of lightweight architectures/backbones for this purpose (HardNet, MobileNetV2, MobileNetV2plus, ShelfNet, FasterSeg, BiseNetV2, SINet, PortraitNet, etc, etc)

Could HRNetV2-W18-Small-v2 be considered SOTA for this task? It achieves mIOU of 76.2% on Cityscapes.

The mIOU of U-HarDNet-70 is 75.9% & it is at the top of paperswithcode SOTA list for real-time

2) Is the CityScapes leaderboard a good place to look for my task at hand? I am looking for human/person segmentation. Should I rather be looking for leaderboards of Coco or PASCAL?

3) Is there "HRNet + OCR + SegFix" option available for HRNetV2-W18-Small-v2?

Thanks for your patience with the newbie questions!

WaterKnight1998 commented 4 years ago

Wow congratulations for the amazing library!

I am looking to implement high-quality semantic segmentation on a mobile device for accurate human segmentation for still images (i.e. non-realtime).

  1. SOTA for light-weight?

I have been researching a lot of lightweight architectures/backbones for this purpose (HardNet, MobileNetV2, MobileNetV2plus, ShelfNet, FasterSeg, BiseNetV2, SINet, PortraitNet, etc, etc)

Could HRNetV2-W18-Small-v2 be considered SOTA for this task? It achieves mIOU of 76.2% on Cityscapes.

The mIOU of U-HarDNet-70 is 75.9% & it is at the top of paperswithcode SOTA list for real-time

  1. Is the CityScapes leaderboard a good place to look for my task at hand? I am looking for human/person segmentation. Should I rather be looking for leaderboards of Coco or PASCAL?
  2. Is there "HRNet + OCR + SegFix" option available for HRNetV2-W18-Small-v2?

Thanks for your patience with the newbie questions!

Which heads have you tried??

InternetMaster1 commented 4 years ago

Sorry, I am new to this. What do you mean by heads?

1) I have tried following library till now, it's based on DeeplabV3+ and Resnet50, but looks heavy for mobile usage. https://github.com/srihari-humbarwadi/person_segmentation_tf2.0

2) I even tried the Deeplabv3plus and Mobilenetv2 combo based on this library. https://github.com/nolanliou/mobile-deeplab-v3-plus/issues/11

That library owner recommended to use a better backbone to reach higher accuracy.

Hence, I am trying to hunt a lightweight segmentation library which would help me achieve higher accuracy on the same training as compared to above mentioned libraries.

Thanks

WaterKnight1998 commented 4 years ago

Sorry, I am new to this. What do you mean by heads?

  1. I have tried following library till now, it's based on DeeplabV3+ and Resnet50, but looks heavy for mobile usage. https://github.com/srihari-humbarwadi/person_segmentation_tf2.0
  2. I even tried the Deeplabv3plus and Mobilenetv2 combo based on this library. nolanliou/mobile-deeplab-v3-plus#11

That library owner recommended to use a better backbone to reach higher accuracy.

Hence, I am trying to hunt a lightweight segmentation library which would help me achieve higher accuracy on the same training as compared to above mentioned libraries.

Thanks

This arch has the backbone included. If you just want the backbone you should go to classification repo

InternetMaster1 commented 4 years ago

No, not looking for just a backbone.

Looking for the best arch+backbone combo for lightweight segmentation with high accuracy.

I was just just curious about HRNet vs HarDNet, and wanted opinion as to which would be considered a better option for my task...