Fortnightly Summary

Implemented U-net head in mmseg, then use U-net head as decoder and ResNet-50 as backbone encoder, run standard VOC training and validation. The preliminary result shows that U-net head, though having more parameters and being more capable to fit the data, performance regarding to MIoU is nevertheless on the short side, when comparing to plain convolution layers with mysterious custom modifications.

Test Report

experiments/moco-unet/superIN1M-linear-30k-2gpus/20200923_154850.log.json