Early Object Detection Networks

Keywords

ver. 1.0.0, early_object_detection_networks

How to build, train and evaluate modern region-based object detectors (R-CNN → Fast R-CNN → Faster R-CNN), their components, and trade-offs

This unit explains region-based object detection: framing detection as classifying and localising region proposals, adapting ImageNet CNNs by fine-tuning, improving localisation with bounding-box regression, sharing convolutional computation with RoI pooling, training with a multi-task loss, and learning proposals with a Region Proposal Network so detection and proposals share a backbone. By the end you can train and evaluate two-stage detectors, explain design choices, and compare speed/accuracy trade-offs across R-CNN, Fast R-CNN and Faster R-CNN.

Full summary

The unit teaches how to go from whole-image classification to detecting and precisely localising every object in an image using region-based detectors. You learn the object detection problem and how detectors are evaluated (IoU-based mean average precision), then follow the R-CNN approach: generate category-independent region proposals, warp each proposal and extract CNN features, and classify proposals with per-class linear SVMs. Because detection datasets are small, you adapt an ImageNet-pretrained convolutional network by domain-specific fine-tuning on warped proposals and justify the label and classifier choices.

You learn to read detection error analyses and fix the dominant failure mode—poor localisation—by adding class-specific bounding-box regression to refine proposal coordinates. To address R-CNN’s inefficiency, the unit introduces RoI pooling so all proposals share one convolutional forward pass, and shows how to collapse multi-stage training into a single-stage optimisation with a joint classification and localisation (multi-task) loss. Practical bottlenecks and speedups are covered: with shared convolutions the fully-connected layers become the bottleneck, so techniques such as truncated SVD and calibrated numbers of proposals are used to trade off accuracy and runtime.

The unit then removes the dependence on external proposal methods by training a small fully convolutional Region Proposal Network (RPN) that predicts objectness scores and box offsets relative to translation-invariant anchor boxes placed on the shared feature map. You study strategies for training the RPN and detector to share one backbone and the costs/benefits of sharing. Finally, you compare R-CNN, Fast R-CNN and Faster R-CNN quantitatively on VOC/COCO and qualitatively in terms of which bottleneck each paper removed, and are prepared to explain why the two-stage cascade remains strong and how it connects to single-shot detectors.

By the end of the unit you can: - State the detection task and compute IoU-based mAP to evaluate detectors. - Describe the R-CNN pipeline (proposals → CNN features → per-class classifiers) and the rationale for pretraining and fine-tuning. - Apply and justify bounding-box regression for localisation improvement. - Implement and explain RoI pooling to share convolutional computation across proposals. - Formulate and train a detector with a multi-task classification/localisation loss in one stage. - Design and train a Region Proposal Network with anchors, and integrate it so RPN and detector share convolutional features. - Compare the accuracy and runtime trade-offs of R-CNN, Fast R-CNN and Faster R-CNN, and identify the bottleneck each removed.

This unit gives the knowledge and practical reasoning needed to build, train and diagnose two-stage region-based detectors and to understand the subsequent move toward one-stage, single-shot detection methods.

Materials

Lecture notesThe unit written out, to read at your own pacever. 1.1.6 · 2026-10-02 01:11:15Open → SlidesThe deck to present fromver. 1.0.13 · 2026-10-02 02:06:30Open →

Source documents

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.
Girshick, Ross. “Fast r-cnn.” Proceedings of the IEEE international conference on computer vision. 2015.
Ren, S., He, K., Girshick, R., & Sun, J. (2015). Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28.

← Residual Convolutional Networks Fast Object Detection Networks →

--- title: "Early Object Detection Networks" keywords: - ver. 1.0.0 - "early_object_detection_networks" format: html: toc: false --- **How to build, train and evaluate modern region-based object detectors (R-CNN → Fast R-CNN → Faster R-CNN), their components, and trade-offs** > This unit explains region-based object detection: framing detection as classifying and localising region proposals, adapting ImageNet CNNs by fine-tuning, improving localisation with bounding-box regression, sharing convolutional computation with RoI pooling, training with a multi-task loss, and learning proposals with a Region Proposal Network so detection and proposals share a backbone. By the end you can train and evaluate two-stage detectors, explain design choices, and compare speed/accuracy trade-offs across R-CNN, Fast R-CNN and Faster R-CNN. ::: {.callout-note collapse="true" title="Full summary"} The unit teaches how to go from whole-image classification to detecting and precisely localising every object in an image using region-based detectors. You learn the object detection problem and how detectors are evaluated (IoU-based mean average precision), then follow the R-CNN approach: generate category-independent region proposals, warp each proposal and extract CNN features, and classify proposals with per-class linear SVMs. Because detection datasets are small, you adapt an ImageNet-pretrained convolutional network by domain-specific fine-tuning on warped proposals and justify the label and classifier choices. You learn to read detection error analyses and fix the dominant failure mode—poor localisation—by adding class-specific bounding-box regression to refine proposal coordinates. To address R-CNN's inefficiency, the unit introduces RoI pooling so all proposals share one convolutional forward pass, and shows how to collapse multi-stage training into a single-stage optimisation with a joint classification and localisation (multi-task) loss. Practical bottlenecks and speedups are covered: with shared convolutions the fully-connected layers become the bottleneck, so techniques such as truncated SVD and calibrated numbers of proposals are used to trade off accuracy and runtime. The unit then removes the dependence on external proposal methods by training a small fully convolutional Region Proposal Network (RPN) that predicts objectness scores and box offsets relative to translation-invariant anchor boxes placed on the shared feature map. You study strategies for training the RPN and detector to share one backbone and the costs/benefits of sharing. Finally, you compare R-CNN, Fast R-CNN and Faster R-CNN quantitatively on VOC/COCO and qualitatively in terms of which bottleneck each paper removed, and are prepared to explain why the two-stage cascade remains strong and how it connects to single-shot detectors. By the end of the unit you can: - State the detection task and compute IoU-based mAP to evaluate detectors. - Describe the R-CNN pipeline (proposals → CNN features → per-class classifiers) and the rationale for pretraining and fine-tuning. - Apply and justify bounding-box regression for localisation improvement. - Implement and explain RoI pooling to share convolutional computation across proposals. - Formulate and train a detector with a multi-task classification/localisation loss in one stage. - Design and train a Region Proposal Network with anchors, and integrate it so RPN and detector share convolutional features. - Compare the accuracy and runtime trade-offs of R-CNN, Fast R-CNN and Faster R-CNN, and identify the bottleneck each removed. This unit gives the knowledge and practical reasoning needed to build, train and diagnose two-stage region-based detectors and to understand the subsequent move toward one-stage, single-shot detection methods. ::: ## Materials ::: {.hg-deck} <a class="hg-card" href="notes.qmd">Lecture notesThe unit written out, to read at your own pacever. 1.1.6 · 2026-10-02 01:11:15Open →</a> <a class="hg-card" href="slides.qmd">SlidesThe deck to present fromver. 1.0.13 · 2026-10-02 02:06:30Open →</a> ::: ## Source documents - 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. - Girshick, Ross. "Fast r-cnn." Proceedings of the IEEE international conference on computer vision. 2015. - Ren, S., He, K., Girshick, R., & Sun, J. (2015). Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28. ::: {.hg-nav} [← Residual Convolutional Networks](../residual_convolutional_networks/index.qmd){.hg-prev} [Fast Object Detection Networks →](../fast_object_detection_networks/index.qmd){.hg-next} :::