baidu-research / NCRF

Cancer metastasis detection with neural conditional random field (NCRF)
Apache License 2.0
757 stars 184 forks source link

downsampling level selection #5

Closed sketchplanet closed 6 years ago

sketchplanet commented 6 years ago

The heatmap generated by mask downsampled in level 8 and the heatmap generated by mask downsampled in level 6 difffers a lot. How to choose the downsampling level when generating the heatmap ? Why cause this? A little confused , Thank you!

yil8 commented 6 years ago

@sketchplanet Generally, speaking smaller level heat map means smaller inference stride, e.g. (level 8 = 256, vs level 6 = 64). Smaller inference stride typically get you higher sensitivity with the sacrifice of more computation. The wining team of camelyon16 used an inference stride of 4, which is computationally infeasible in my current implementation for obtaining the heat map.

sketchplanet commented 6 years ago

@yil8 Previously I generate heatmap in level 8(extracting patch without overlap). I didn't realize the huge difference bwtween different strides until try your code. Maybe this is what makes me stuck in recent months. Thank you.