Distance study

our goal is to find adaptive loss function. (able to measure distribution concept)

TODO List

• L2 Loss
• L1 Loss
• BCE Loss
• EM divergence
• KL divergence
• JSD divergence

Setting

• 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]