Skip to content

fix for humanbody vs. others#3050

Merged
MMathisLab merged 46 commits intomainfrom
mwm/humanbody
Jul 8, 2025
Merged

fix for humanbody vs. others#3050
MMathisLab merged 46 commits intomainfrom
mwm/humanbody

Conversation

@MMathisLab
Copy link
Copy Markdown
Member

@MMathisLab MMathisLab commented Jul 6, 2025

  • There was a bug in rc11 that causes the non humanbody SA models to attempt to use the torchvision COCO detector vs. the custom ones. This fixes that error.

  • SA_TVM is not just fully functional, but it's a WIP for @MMathisLab

  • It also required a fix in dlclibrary, which needs merged before this: Update modelzoo_urls_pytorch.yaml DLClibrary#41

  • It also cleans up the GUI a bit to remove the un-needed video type button. The code automatically handles this anyhow.

  • It has two detectors for humans now as well: "fasterrcnn_mobilenet_v3_large_fpn", "fasterrcnn_resnet50_fpn_v2"

  • I did some speed profiling locally on my M2 Apple Chip macbook; but only using the CPU on a short video (so just consider relative perf. speed, not as top-end)

Results:

  • All 9 superanimal_quadruped model + detector combinations on jasper-short.mov were successful.
  • All humanbody passed
  • All bird passed
  • The fasterrcnn_resnet50_fpn_v2 detector now works for all backbones (no failures, no fallback to torchvision COCO weights).
  • No errors, and all output files were generated as expected.

Short Results

Dataset Model Detector Status Time (s)
quadruped hrnet_w32 fasterrcnn_mobilenet_v3_large_fpn 2.74
quadruped hrnet_w32 fasterrcnn_resnet50_fpn_v2 11.49
quadruped hrnet_w32 ssdlite 6.61
quadruped resnet_50 fasterrcnn_mobilenet_v3_large_fpn 2.61
quadruped resnet_50 fasterrcnn_resnet50_fpn_v2 10.31
quadruped resnet_50 ssdlite 3.15
quadruped rtmpose_s fasterrcnn_mobilenet_v3_large_fpn 1.87
quadruped rtmpose_s fasterrcnn_resnet50_fpn_v2 9.58
quadruped rtmpose_s ssdlite 2.56
humanbody rtmpose_x fasterrcnn_mobilenet_v3_large_fpn 14.22
humanbody rtmpose_x fasterrcnn_resnet50_fpn_v2 13.50
bird resnet_50 fasterrcnn_mobilenet_v3_large_fpn 2.74
bird resnet_50 ssdlite 1.68

Extended Tests Passing

GIST for the testscript: https://gist.github.com/MMathisLab/6a4ba42af34b43b2d9b2be4040ee02cf

📊 Overall Summary

  • Total Tests: 13 combinations
  • ✅ Successful: 13
  • ❌ Failed: 0
  • Success Rate: 100.0%

⏱️ Performance Statistics

  • Average Time: 12.88s
  • Fastest: 2.38s
  • Slowest: 26.49s

🦄 SuperAnimal Quadruped

Success Rate: 9/9 (100.0%)

Models & Detectors

Model Detector Status Time
hrnet_w32 fasterrcnn_mobilenet_v3_large_fpn 8.36s
hrnet_w32 fasterrcnn_resnet50_fpn_v2 26.49s
hrnet_w32 ssdlite 18.17s
resnet_50 fasterrcnn_mobilenet_v3_large_fpn 6.02s
resnet_50 fasterrcnn_resnet50_fpn_v2 23.18s
resnet_50 ssdlite 6.83s
rtmpose_s fasterrcnn_mobilenet_v3_large_fpn 4.39s
rtmpose_s fasterrcnn_resnet50_fpn_v2 19.44s
rtmpose_s ssdlite 3.67s

Average Time: 12.95s


👤 SuperAnimal Humanbody

Success Rate: 2/2 (100.0%)

Models & Detectors

Model Detector Status Time
rtmpose_x fasterrcnn_mobilenet_v3_large_fpn 21.64s
rtmpose_x ssdlite_mobilenet_v3_large 22.55s

Average Time: 22.09s


🐦 SuperAnimal Bird

Success Rate: 2/2 (100.0%)

Models & Detectors

Model Detector Status Time
resnet_50 fasterrcnn_mobilenet_v3_large_fpn 4.34s
resnet_50 ssdlite 2.38s

Average Time: 3.36s


🔧 Technical Details

Output Files

All results saved to: /Users/mackenzie/Desktop/testing/comprehensive_superanimal_tests/

Test Videos

  • Quadruped: jasper-short.mov (5 frames)
  • Humanbody: bio-very-short.mov (19 frames)
  • Bird: bird.mov (14 frames)

@MMathisLab MMathisLab requested a review from AlexEMG July 6, 2025 20:41
@AlexEMG AlexEMG requested a review from yeshaokai July 6, 2025 21:03
@MMathisLab MMathisLab requested review from AlexEMG and removed request for yeshaokai July 7, 2025 06:47
-remove debug print statement
* all working

* Update fasterRCNN.py

- remove the extra debug statements
@MMathisLab
Copy link
Copy Markdown
Member Author

To be clear, this fixes the failing superanimal models except TVM + mobilenet; everything else works.

@MMathisLab MMathisLab force-pushed the mwm/humanbody branch 2 times, most recently from 1224a57 to 8af154b Compare July 7, 2025 16:08
@MMathisLab MMathisLab merged commit 5374737 into main Jul 8, 2025
5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants