Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation

Inside the paper

The problem

Progress on object detection had slowed, and a strong image classifier did not by itself say where objects were. Detection datasets also offered much less labeled data than large classification datasets.

The idea

R-CNN applies a convolutional network to proposed image regions, then classifies those regions. The network is first trained for classification and adapted to the detection task.

How it works

• A region proposal method supplies candidate object locations without requiring a separate proposal generator for each class.
• Each candidate region is warped to a fixed size and passed through the network to obtain a feature vector.
• The authors fine-tune the pretrained network on labeled detection regions, then train class-specific linear support vector machines on the resulting features.
• A learned bounding-box regressor refines predicted locations, and overlapping detections are filtered during the final detection stage.

What it showed

The paper reports substantial improvements over previous approaches on PASCAL VOC detection and compares its detector with a sliding-window system on the ILSVRC detection dataset. It also tests the region features for semantic segmentation.

Read it for

Figure 1 and Section 2 lay out the pipeline; Section 3 examines what the learned features contribute. Section 4 is a useful second pass if segmentation is your interest, while implementation details can wait.

Written in our own words from the paper itself. A draft the editors have not yet reviewed.