Arxiv 2018

[arxiv 1710.10710] On Pre-Trained Image Features and Synthetic Images for Deep Learning [PDF] [notes]

Stefan Hinterstoisser, Vincent Lepetit, Paul Wohlhart, Kurt Konolige

read 04/25/2017

Objective

• show advantage of freezing feature extractor when training object detection on synthetic data
• compare full training and partial on RFCN and MaskR-RCNN

Synthesis

• Freeze feature extractor and train only the rest of the detection pipeline
• Use inception-resnet and Resnet 101 as base
• Show close-to-real results

Results

Show that blurring gives simple but efficient boost in precision. Blurring is operated using a Gaussian kernel including the boundaries to blur the object with the background

Show that pairs of real/synthetic images have small difference in feature space using frozen layers, otherwise show that when training on synthetic this difference is much larger

Show that freezing the whole feature extractor is more efficient then freezing only the bottom layers (the more is frozen, better it is). However, this finding is contradicted by arxiv 1804.06516 which evaluates on car detection and gets significant drop in performance when freezing the feature extractor.