search

sabarim / STEm-Seg

This repository contains the official implementation of the paper "STEm-Seg: Spatio-temporal Embeddings for Instance Segmentation in Videos"
153 stars 23 forks source link

code question #1

Closed cymqqqq closed 4 years ago

cymqqqq commented 4 years ago

can U explain 👍 self.conv_16 = nn.Conv3d(inter_channels[0] + inter_channels[1], inter_channels[1], 1, bias=False) what's the meaning of inter_channels[0] + inter_channels[1]

sabarim commented 4 years ago

I hope that you are referring to embedding_decoder.py file. Here, the input to this convolutional layer is a concatenated feature map from self.block_32x and self.block_16x, where the output channels of these individual blocks are inter_channels[0] and inter_channels[1] respectively. The concatenation would make the number of input channels to the self.conv_16 layer to be inter_channels[0] + inter_channels[1]. Please refer to the function "forward" for a better understanding on this.

Also, the inter_channels list itself comes from the config file which is passed to the python program as a parameter. Please check stemseg/config/defaults.yaml for more details.

cymqqqq commented 4 years ago

thank you