RCNNConfig

Configuration for R-CNN.

R-CNN (Girshick et al., CVPR 2014) applies a CNN independently to each region proposal (warped to a fixed roi_size × roi_size crop), then feeds the flattened pool5 features through two FC layers before the classification and bounding-box regression heads.

Original paper uses AlexNet (5 conv + 3 FC) as the backbone. The effective output of pool5 for a 227 × 227 input is 6 × 6 × 256 = 9 216.

Args: num_classes: Number of foreground object classes. Background is automatically added as class 0, giving (num_classes + 1) output logits. in_channels: Number of input image channels (3 for RGB). context_pad: Pixels of image context kept around each proposal in the warped frame (Appendix A's $p$). 0 warps the tight box. roi_size: Each region proposal is warped to this square size before being forwarded through the backbone. Original paper: 227. dropout: Dropout probability applied after fc6 and fc7. score_thresh: Minimum class-score threshold applied during post-processing (after softmax). Boxes whose max class score is below this value are discarded. nms_thresh: IoU threshold for per-class NMS at inference time. max_detections: Maximum number of detections returned per image after NMS.

Note:

Training — two of the paper's three stages. Passing targets to forward trains stage 1 (CNN fine-tuning, softmax cross-entropy over proposals at IoU >= fg_iou_thresh) and stage 3 (the class-specific bounding-box regressors on pool5 features, restricted to proposals at IoU > bbox_reg_iou_thresh per Appendix C).

Stage 2 — the per-class linear SVMs that replace the softmax at test time — is not part of the model. It is an offline pass over features cached from a frozen network, with positives drawn only from ground-truth boxes, negatives at IoU < 0.3, and hard negative mining across the whole dataset; the mining is what makes it work, and it is a pipeline rather than a layer. lucid.models._utils._common.mine_hard_negatives implements that loop over cached features — score the pool, keep the highest-scoring mistakes, refit, repeat — on top of lucid.models._utils._common.fit_linear_svm. Extracting the features from a frozen network over a whole dataset, and scoring with the resulting weights instead of the softmax, remain the caller's pipeline.