Implement newer FAST Corner Detector #460
Conversation
This comment was marked as outdated.
This comment was marked as outdated.
|
Nvm, |
- Replace free functions with FastDetector struct and methods - Update benchmarks, tests, and example to use FastDetector - Improve mask handling and keypoint extraction logic
- Rename `corner_fast` to `compute_corner_response` - Rename `get_keypoints` to `extract_keypoints` - Add and expand doc comments for public methods and struct fields - Refactor internal logic for clarity and parallelism - Update examples and benchmarks to use new method names
There was a problem hiding this comment.
Pull Request Overview
This PR implements a newer FAST (Features from Accelerated Segment Test) corner detector, replacing the previous function-based implementation with a stateful FastDetector struct. The new implementation includes corner response computation, non-maximum suppression, and configurable minimum distance between keypoints.
Key changes:
- Replaced the simple
fast_feature_detectorfunction with a comprehensiveFastDetectorstruct - Added corner response computation and keypoint extraction as separate steps
- Introduced minimum distance filtering to reduce clustering of detected keypoints
Reviewed Changes
Copilot reviewed 5 out of 5 changed files in this pull request and generated 3 comments.
Show a summary per file
| File | Description |
|---|---|
| examples/fast_detector/src/main.rs | Updated to use new FastDetector API with additional parameters and separate computation steps |
| examples/fast_detector/README.md | Updated documentation to reflect new command-line interface and parameters |
| examples/fast_detector/Cargo.toml | Removed unnecessary "gstreamer" feature dependency |
| crates/kornia-imgproc/src/features/fast.rs | Complete rewrite implementing stateful FastDetector with corner response computation and keypoint extraction |
| crates/kornia-imgproc/benches/bench_features.rs | Updated benchmark to use new FastDetector API and added detector clearing |
Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.
| if consecutive_count == n { | ||
| curr_response = 0; | ||
| circle_intensities.iter().for_each(|m| { | ||
| curr_response = curr_response.saturating_add(m.saturating_sub(curr_pixel)); |
There was a problem hiding this comment.
The corner response calculation is incorrect. When m < curr_pixel, m.saturating_sub(curr_pixel) returns 0, which doesn't contribute to the response. For darker pixels, the calculation should use curr_pixel.saturating_sub(*m) to get the actual intensity difference.
| curr_response = curr_response.saturating_add(m.saturating_sub(curr_pixel)); | |
| curr_response = curr_response.saturating_add((i16::from(*m) - i16::from(curr_pixel)).abs() as u8); |
Copilot uses AI. Check for mistakes.
| (yy as isize - y as isize) | ||
| .abs() | ||
| .max((xx as isize - x as isize).abs()) | ||
| < min_distance as isize |
There was a problem hiding this comment.
The distance calculation uses Chebyshev distance (max of absolute differences) but the condition should be <= min_distance instead of < min_distance to properly suppress points exactly at the minimum distance.
| (yy as isize - y as isize) | |
| .abs() | |
| .max((xx as isize - x as isize).abs()) | |
| < min_distance as isize | |
| <= min_distance as isize |
Copilot uses AI. Check for mistakes.
$ cargo run -p fast_detector --release -- --image-path "path/to/image.png" --min-distance 10 --threshold 30Benchmark
TODOs