module

R-CNN

3 members

lucid.models.vision.rcnn

R-CNN — object detection (Girshick et al., CVPR 2014).

Paper: "Rich feature hierarchies for accurate object detection and semantic segmentation"

Object Detection

Paper

Girshick, Ross, et al. "Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2014, pp. 580–587.

Overview

R-CNN (Regions with CNN features) is the first deep-learning detector to demonstrate that ImageNet-pretrained CNN features transfer to localisation. The pipeline is deliberately modular:

Region proposals. An external method (selective search in the original paper) yields $\sim 2000$ class-agnostic candidate windows per image.
CNN feature extraction. Each proposal is warped to a fixed $227 \times 227$ patch and forwarded through an ImageNet-pretrained CNN (AlexNet). The pool5 activations (effectively $6 \times 6 \times 256 = 9216$ features) are flattened and passed through two fully-connected layers.
Per-class linear SVMs score each region; a class-specific bounding-box regressor refines the proposal coordinates with a parameterised $(t_x, t_y, t_w, t_h)$ offset.

The cost is dominated by redundant CNN forward passes — one per proposal — because no feature sharing across overlapping regions exists. Despite that inefficiency, R-CNN jumped mean average precision on PASCAL VOC 2012 from $\sim 35\%$ (DPM) to $53.3\%$ , making CNNs the dominant detection paradigm and motivating every successor in the family (Fast R-CNN, Faster R-CNN, Mask R-CNN).

Classes

dataRCNNConfig1 methods

Configuration for R-CNN.

classRCNNForObjectDetection3 methods

R-CNN object detector (Girshick et al., CVPR 2014).

Functions

fnrcnn→ RCNNForObjectDetection

R-CNN with AlexNet backbone (Girshick et al., CVPR 2014).