This repository implemented for experiment of skip-connection's power.
I experimented using Dense-U-Net which consist of dense block and bottleneck block.
Dense block and bottleneck block are like under picture.
Dense blocks were used in U-Net main block. and bottleneck blocks were used in skip-connection.
So main model's appearance is U-Net. and I tried 4 types networks.
Base model used only base skip-connection looks like a u-net. others used additional up or down
When I merged the skip connection and up stream, I used concatenation and additional conv block.
I used mph_dataset.
It provided 2D, 3D points and bounding box about hands. and I used 2D, 3D points.
And I used 45,581 for training, 13,023 for validation and 6,511 for test.
For 2D estimation, I used heatmap format. and 3D Estimation used relative points which origin to palm.
and used Focal Loss for 2D points and MSE Loss for 3D Points estimation.
I measured performance of networks using PCK(PCK 2D, PCK 3D).
PCK 2D measure amount of pixel between predicted and ground-truth.
PCK 3D measure percentage of distance between prediction and ground-truth about base distance like j1-parm distance.
and under pictures saw the result of model training. validation loss in training step. and PCK 2D, PCK 3D in test dataset.
What I expected was using all skip-connection will have best performance.
but actually result showed the using additional under skip-connection has best performance.
- Inference code implement.
- training code using argment