[code] optimize sliding window trick to selective search

[clay-arras]

Prerequisites

The current implementation uses a sliding window approach for inference, which is relatively slow. You should have a functioning CNN binary classifier and a crop_to_bbox utility already implemented.

Motivation

To improve inference speed, we need to transition from a computationally expensive sliding window to a Selective Search region proposal method. Region proposal is just a CV heuristic that groups pixels with similar color and texture to generate a list of rectangles that COULD contain a valid object. This effectively turns our object localization detector into a two-pass model.

Overview

Refactor the existing localization logic to utilize Selective Search for region proposals, followed by CNN classification for detection.

Implementation Checklist

• Create Region Proposal Class: Implement a separate class for the region proposal portion using OpenCV’s Selective Search. Link below on what that might look like.

  • Note: Since the algorithm outputs a list of rectangle coordinates and the CNN requires fixed-size images, ensure the output regions are cropped and resized correctly; this logic should exist already in the crop_to_bbox method in the PascalVOCDataset
  • I recommend extracting the cropping and resizing logic from crop_to_bbox into a standalone utility function, just for better separation of responsibility.

• Encapsulate Object Localizer: Create a top-level class for the object localizer that inherits from torch.nn.Module. This class should encapsulate both the selective search (for region proposals) and the binary classifier (for detection). create a new file for this.

Reference

For implementation details on Selective Search in C++, see: LearnOpenCV - Selective Search for Object Detection

[clay-arras]

correction: selective search isn't a RPN, it's a region proposal algorithm, updated the issue