5,337
edits
Line 663: | Line 663: | ||
===Homework=== | ===Homework=== | ||
Read OHEM paper. | Read [https://arxiv.org/pdf/1604.03540.pdf OHEM paper]. | ||
Summary of online hard example mining: | |||
Previously hard mining involved the following two steps | |||
# Freeze the model and run it on 10-100s to find hard examples. | |||
# Train on hard examples. | |||
They instead propose finding hard-examples per mini-batch. | |||
This is possible because there are thousands of RoIs within each image. | |||
# Run a mini-batch through the CNN feature extractor. | |||
# Do forward-pass on all RoIs. | |||
# Sort RoIs by loss and take the top <math>B/N</math> examples. | |||
#* Filter duplicates using NMS. | |||
# Backwards pass only on the top <math>B/N</math> examples. | |||
==Will be on the exam== | ==Will be on the exam== |