our goal is to find adaptive loss function. (able to measure distribution concept)
- L2 Loss
- L1 Loss
- BCE Loss
- EM divergence
- KL divergence
- JSD divergence
- dataset: SKU110k dataset
in main.py 60 lines
images = datas[0]
boxes = datas[1]
labels = datas[2]
locations = datas[3]
cnt = datas[4]
map = datas[5] (Gaussian)
HD
HD (Improved)
L1
BCE
JSD
HD + BCE
- model : depth-wise separable convolutional unet
- output of model : [B, 100, 100], [B]