LECTURE 23

Summary

The lecture explains the development of object-detection methods from R-CNN to Fast R-CNN and Faster R-CNN, then discusses adaptations for detecting objects in SAR imagery.

From R-CNN to Fast R-CNN

The original R-CNN pipeline involves several costly, separate steps:

1. Generate region proposals, for example, using selective search.
2. Resize each proposed region and run it through a CNN independently.
3. Classify each region with an SVM.
4. Refine the detected object’s bounding box.

Running the CNN separately for every proposal repeats much of the same computation. Training is also divided across multiple stages.

Fast R-CNN reduces this redundancy:

• Process the full image once: A CNN produces a convolutional feature map for the image.
• Reuse the features: Region proposals, still generated separately (for example, with selective search), are projected onto the feature map.
• Use ROI pooling: Each proposed region is divided into fixed spatial bins and pooled to produce a fixed-size feature vector.
• Predict jointly: Fully connected layers produce both object-class scores, including a background class, and refined bounding-box coordinates.

This approach shares convolutional computation across proposals and combines classification and localization in one network. Its multitask loss combines a classification loss (log loss) with a bounding-box regression loss (a robust L1 loss). Regression is applied only to foreground objects, not to background proposals.

Faster R-CNN

Fast R-CNN still relies on an external proposal-generation method, which can be slow and is not learned jointly with the detector. Faster R-CNN addresses this limitation by adding a Region Proposal Network (RPN):

• The RPN operates on the CNN feature map.
• It proposes bounding boxes and assigns objectness scores to them.
• Its convolutional features are shared with the detector.

The key change is that proposal generation becomes part of the learned architecture rather than relying on a separate method such as selective search.

Adapting Detection to SAR Imagery

The lecture presents two approaches that account for SAR data characteristics and limited annotation.

SAR Ship Detection with Preprocessing

Ships may occupy only a few pixels in a larger scene, particularly at medium resolution. One approach narrows the search before applying an object detector:

1. Use preprocessing, such as morphology and preclassification, to identify smaller candidate patches likely to contain ships.
2. Analyze candidate regions—for example, with HOG or another method—to estimate the dominant object orientation.
3. Rotate or otherwise align the candidate patches to a chosen orientation.
4. Apply an R-CNN-family detector to the processed regions.

The aim is to make detection more manageable and improve detector performance.

Semi-Supervised SAR Target Detection

Training CNN-based SAR detectors typically requires many labeled images, but SAR annotation is costly and requires expertise. The lecture describes a strategy for making use of unlabeled data:

1. Train a detector on a labeled SAR image type (for example, one band), using its encoder and detection heads.
2. Use a frozen encoder from that labeled source alongside a trainable encoder for a second SAR image type with little or no labeled data.
3. Encourage the trainable encoder’s features to align with the source encoder’s feature distribution, assuming the two data types depict similar kinds of areas.
4. Train the second-type detector with any available detection labels.
5. Optionally, add a reconstruction decoder and reconstruction loss to further train the encoder.

The same general idea could transfer knowledge from labeled optical imagery to SAR imagery, although the example discussed focuses on transferring between SAR image types.

Speakers and Sources

• Speaker: One lecturer, whose name is not given in the subtitles.
• Sources and methods discussed: R-CNN, Fast R-CNN, Faster R-CNN, selective search, ROI pooling, the Region Proposal Network, HOG-based orientation analysis, and a SAR ship-detection paper cited in the lecture without identifying details. No other speakers are identifiable.