search

karfly / learnable-triangulation-pytorch

This repository is an official PyTorch implementation of the paper "Learnable Triangulation of Human Pose" (ICCV 2019, oral). Proposed method archives state-of-the-art results in multi-view 3D human pose estimation!
MIT License
1.09k stars 181 forks source link

Evaluation of Volumetric Model #123

Open sunwonlikeyou opened 3 years ago

sunwonlikeyou commented 3 years ago

I simulated pretrained volumetric model with pretrained backbone. and i got these result. 'S11': {'Average': 158.4710789766407, 'Directions': 13.81251708984375, 'Discussion': 20.87702763545049, 'Eating': 107.31265129841549, 'Greeting': 104.50263323102679, 'Phoning': 249.43584526909723, 'Posing': 33.8868872601053, 'Purchases': 160.68772977941177, 'Sitting': 87.17900390625, 'SittingDown': 264.5645211641906, 'Smoking': 240.20394674161585, 'TakingPhoto': 190.77657277960526, 'Waiting': 58.71925354003906, 'Walking': 307.15990271935095, 'WalkingDog': 243.03764997209822, 'WalkingTogether': 201.88863716873468}

and also i checked heatmap from backbone in volumetric model image

but heatmap from backbone in algebraic model is image

is there problem or mistake on evaluation??

shrubb commented 3 years ago

Looks like you caught a bug. These heatmaps seem unnatural to me, as if the network is underfit or overfit. For example, I believe there shouldn't be such checkerboard patterns. You probably messed up .requires_grad, .eval() vs .train() or something like that.

sunwonlikeyou commented 3 years ago

Backbones were all eval() and also all tensors' requires_grad were also False. And That plottings are heatmap of 'Directions-1' which shows good quantitative result but i can check poor heatmap result. And also I checked Loading pretrained weights for backbone and whole model. can you show me your log files??

shrubb commented 3 years ago

Hi, unfortunately I don't have logs :( I still bet you have a bug somewhere in the backbone. I suggest that you try checking activations and gradients layer-wise.