Region proposal

As explained in Math and Architectures of Deep Learning

A candidate image region proposed as possibly containing an object for later classification and localization.

A region proposal, or region of interest (RoI), is a candidate image region that may contain an object. Two-stage detectors use proposals to separate “where might an object be?” from classification and bounding-box refinement. In R-CNN, selective search produces about 2,000 candidate regions per image; each is labeled positive or negative during training from its overlap with ground-truth boxes. section 11.3.1 section 11.4.4

Proposal generation moved from a separate computer-vision algorithm to a learned network. Faster R-CNN uses a region proposal network (RPN) on convolutional feature maps to predict candidate RoIs across a range of scales and aspect ratios. The RPN and the downstream Fast R-CNN detector share convolutional layers, avoiding separate backbones. section 11.3.3 section 11.4 section 11.4.4

From a proposed region to a detection

A backbone reduces an input image to a convolutional feature map. For an RoI of (0, 0, 256, 256) with subsampling factor 16, the corresponding feature-map region is (0, 0, 16, 16). RoI pooling converts regions of differing sizes into fixed-sized features, which classifier and regressor layers use to predict an object class or background and bounding-box coordinates. section 11.4.3 section 11.4

During Faster R-CNN training, proposals connect the RPN and Fast R-CNN stages. Alternate optimization repeatedly trains the RPN and detector in turn. Approximate joint training instead combines their losses in one network, treating generated proposals as fixed during the detector update; this is faster but ignores a derivative through the proposals. section 11.4.4

Figure 11.18 A conv feature map with two regions of interest of different sizes. RoI pooling extracts a fixed-sized output feature map (2 × 2 in this image) from each of the RoIs in a single pass via max pooling. This enables us to extract fixed-sized representative feature vectors for each RoI, which are then fed into further classifier and regressor layers for classification and localization.

Compare differently sized RoIs on one feature map and see how RoI pooling produces equally sized representations for the detection head. section 11.4.3

How Math and Architectures of Deep Learning covers it

How deep it goes

Explains proposal generation, shared feature maps, RoI pooling, and two training strategies, while deferring selective-search details.

Examples

Uses roughly 2,000 selective-search proposals and maps a 256 × 256 RoI to a 16 × 16 feature-map region.

What you can do after

Evaluate/understand how candidate regions become class scores and refined bounding boxes in a two-stage detector.

How it connects

Connects region proposals to R-CNN, Faster R-CNN, region proposal networks, RoI pooling, convolutional feature maps, and bounding-box regression.

liveBookmanning.com

Read Math and Architectures of Deep Learning on liveBook Get the book at manning.com

Also explained in

• 

  liveBookmanning.com
  Inside Deep Learning: Math, Algorithms, ModelsAn object-detection approach that first proposes candidate regions likely to contain objects before classifying them.

Compare how 2 books explain region proposal →

Related concepts

• Region Proposal Networkproduced by · A Region Proposal Network is a neural-network module that generates candidate object regions from image features for a two-stage detector.
• Faster R-CNNhas part · Faster R-CNN is an object-detection architecture that uses a learned region proposal network and detection heads to localize and classify objects in images.

This explainer was generated by Manning's knowledge graph from the text of Math and Architectures of Deep Learning. The section links point to the parts of the book it draws on.