NoonBody SemanticSegmentation

1 minute read

Task: Semantic Segmentation of Body

  • labels are come from DensePose.
  • Using these body labels, inference body parts.

Training Records

mIoU img_size Model bs epoch opt sch lr aug Β  Β 
0.4348 384,512 mobilenet_v2 + Unet 8 15 AdamW CosineAnnealingWarmRestarts 1e-2 Resize, CenterCrop Β  Β 
0.4968 768, 1024 timm-mobilenetv3_large_100 + Unet 4 15 AdamW CosineAnnealingWarmRestarts 1e-2 Resize, CenterCrop Β  Β 
0.5347 384,512 timm-tf_efficientnet_lite4 + Unet 8 15 AdamW CosineAnnealingWarmRestarts 1e-2 Resize, CenterCrop Β  Β 
0.6226 384,512 timm-efficientnet-b3 + Unet 8 25 RAdam ReduceLROnPlateau(patience=3, factor=0.5) 1e-3 Resize, CenterCrop, CLAHE, RandomBrightnessContrast Β  Β 
0.6285 384,512 timm-efficientnet-b3 + Unet 8 25 RAdam ReduceLROnPlateau(patience=3, factor=0.5) (10e)+ CosineAnnealingWarmRestarts(15e) 1e-3 Resize, CenterCrop, CLAHE, RandomBrightnessContrast Β  Β 
0.6408 384,512 efficientnet-b3 + Unet 8 40 RAdam ReduceLROnPlateau(patience=3, factor=0.5) 1e-3 Resize, CenterCrop, CLAHE, RandomBrightnessContrast Β  Β 
0.6461 384,512 resnet34+ DeepLabV3 16 50 RAdam ReduceLROnPlateau(patience=3, factor=0.5) 1e-2 Resize, CenterCrop, CLAHE, RandomBrightnessContrast Β  Β 
0.6653 384,512 vgg16_bn + Unet 8 40 RAdam ReduceLROnPlateau(patience=3, factor=0.5) 1e-3 Resize, CenterCrop, CLAHE, RandomBrightnessContrast Β  Β 

Trouble Shooting

  • body segmentation label듀이 intμ˜€κΈ° λ•Œλ¬Έμ— cv2.resizeμ—μ„œ error: (-215:Assertion failed) func != 0 in function 'resize'λΌλŠ” μ—λŸ¬κ°€ λ°œμƒν–ˆλ‹€.
    • interpolation 방식을 λ°”κΏ”μ£Όλ©΄ int 인 array도 resizeκ°€ κ°€λŠ₯ν•˜λ‹€.
    • resize interpolation λ°©μ‹μ˜ 차이기 λ•Œλ¬Έμ— λ°œμƒν•˜λŠ” μ—λŸ¬λ‹€.

Review

  • Training Dataset was noisy.
    • Mislabeled
  • Validation Dataset also noisy.
    • Mislabeled also.

Appendix

Reference

https://github.com/facebookresearch/detectron2/blob/main/projects/DensePose/doc/GETTING_STARTED.md

https://kharshit.github.io/blog/2019/08/09/quick-intro-to-semantic-segmentation

https://theaisummer.com/unet-architectures/

http://imlab.postech.ac.kr/dkim/class/csed514_2020s/fcn%20to%20deeplabv3.pdf

https://github.com/qubvel/segmentation_models.pytorch/blob/master/docs/encoders.rst

https://gaussian37.github.io/vision-segmentation-miou/

https://www.kaggle.com/awsaf49/pytorch-sartorius-unet-strikes-back-train#%F0%9F%91%80-Validation-Function

cv2.resize: https://stackoverflow.com/questions/55428929/error-while-resizing-image-error-215assertion-failed-func-0-in-functio

Leave a comment