LECTURE 22

Summary

This lecture introduces object detection and explains the original R-CNN approach, with a focus on how it could be used for SAR imagery. The next lecture is expected to cover how Fast R-CNN and Faster R-CNN address the original method’s limitations.

Object Detection: Goal and Applications

Object detection identifies where objects are and what categories they belong to. Its output includes a class label and a bounding box for each detected object.

Unlike object detection:

• Image classification assigns a label to an entire image.
• Semantic segmentation assigns a label to each pixel.

Detection can be a natural choice when the goal is to locate specific targets—such as ships—rather than label the many surrounding pixels. Earth-observation applications include ship, building, and vehicle detection. The lecture notes that high-resolution vehicle detection may also matter for security and defense.

Detection presents challenges such as objects appearing at different scales, clutter and occlusion, and objects overlapping one another.

From Sliding Windows to R-CNN

A simple detection approach would slide windows across an image, extract features from each window, and classify each one. This can be computationally expensive: it examines many locations and window sizes, even though most image regions are background.

R-CNN improves on this basic idea by separating the process into stages:

1. Generate region proposals. Identify a smaller set of image regions that might contain objects, rather than exhaustively testing every window.
2. Resize each proposal to a fixed size. The regions have different dimensions, but a standard CNN expects fixed-size inputs.
3. Extract CNN features. Process each resized proposal through a CNN to obtain a feature representation.
4. Classify each proposal. Determine whether it contains one of the object categories or belongs to the background.
5. Refine the bounding box. A proposal’s initial box may not fit the object precisely, so bounding-box regression can be used to adjust it.

The lecture compares proposal generation to shortlisting applicants: an initial stage reduces a large set to candidates for further assessment. It is distinct from the later stage that assigns a specific class or rejects a proposal as background.

Region Proposals and Features

• The original R-CNN used selective search to generate region proposals. This method considers image properties such as color, texture, and size; it is not itself a deep-learning method.
• The original system used about 2,000 proposals per image—far fewer than an exhaustive sliding-window search, but still a substantial number.
• Proposal generation should have high recall, so likely objects are included, while keeping computational cost moderate. Many proposals will ultimately be classified as background.
• Since proposals have varying sizes and aspect ratios, R-CNN warps each one to a fixed resolution before feeding it to the CNN. This makes standard CNN processing possible but can distort the region and lose information.
• A generic pre-trained CNN may not be well suited to cropped proposals, which differ in focus and scale from the full images it was trained on. R-CNN therefore fine-tunes the CNN on cropped region proposals to better adapt its features to the task.

Classification and Training

• In the original R-CNN, separate SVM classifiers were trained for the object classes using the CNN features.
• For a given class, a region tightly enclosing an object is treated as a positive example; a background region is a negative example.
• Regions that partially overlap an object are ambiguous. Their treatment can be determined using an intersection over union (IoU) threshold.
• CNN fine-tuning and SVM training are separate stages. The lecturer notes that the original design was modular and that SVMs were still common in computer vision during the transition to CNN-based methods.

Limitations Motivating Later Methods

R-CNN improved on exhaustive search, but it has important costs:

• CNN computation is repeated for every proposal—potentially thousands of times per image.
• Region proposals are generated by a separate, non-deep-learning process.
• Training is multi-stage, with proposal generation, CNN fine-tuning, and SVM classification handled separately.

Fast R-CNN and Faster R-CNN are introduced as later approaches intended to address these inefficiencies.

Speakers and Sources

• One unnamed lecturer/instructor delivers the lecture.
• No other identifiable speakers or sources are featured.