Computer Vision Algorithms: Classical vs Deep Learning

Two controlled studies comparing interpretable classical computer-vision algorithms with pretrained deep models:

1. Corner detection and matching: Harris + SIFT versus SuperPoint on HPatches.
2. Stereo depth estimation: SAD block matching versus RAFT-Stereo on KITTI Stereo 2015.

The classical Harris and SAD algorithms are implemented from scratch. SuperPoint inference is reconstructed around the authors’ pretrained weights, while RAFT-Stereo is used through its official external repository. Dataset files, checkpoints, and third-party source code are deliberately excluded.

Study 1 — Harris vs SuperPoint

The experiment evaluated 20 HPatches sequences—10 illumination and 10 viewpoint sequences—with two target images per sequence, producing 40 image pairs.

Condition	Metric	Harris + SIFT	SuperPoint
Illumination	Repeatability ↑	0.527	0.604
Illumination	Matching score ↑	0.771	0.813
Illumination	Homography error ↓	7.872 px	0.646 px
Viewpoint	Repeatability ↑	0.553	0.409
Viewpoint	Matching score ↑	0.640	0.585
Viewpoint	Homography error ↓	189.498 px	61.607 px

The main finding is deliberately nuanced: SuperPoint produced far more reliable homography estimates, but Harris retained higher raw repeatability and matching score on the selected viewpoint subset. No single metric establishes universal dominance.

Implemented components

• Sobel gradients and structure-tensor construction
• Gaussian smoothing and Harris response
• Thresholding, greedy non-maximum suppression, and top-N selection
• Synthetic validation on a checkerboard and known four-corner square
• SuperPoint architecture and pretrained-weight inference
• SIFT and SuperPoint descriptor matching
• Repeatability, matching score, and RANSAC homography error
• Illumination, viewpoint, and low-texture failure analysis

Study 2 — SAD vs RAFT-Stereo

Both methods were evaluated on the same 40 KITTI Stereo 2015 training pairs.

Metric	SAD	RAFT-Stereo
Mean EPE ↓	16.872 px	0.589 px
Mean D1 ↓	51.122%	1.633%
Mean depth RMSE ↓	115.403 m	1.906 m
Valid prediction ratio ↑	0.619	1.000
Measured runtime	0.656 s	0.659 s

Runtime is reported for completeness, but it is not a hardware-normalised comparison: SAD ran on CPU, while RAFT-Stereo used GPU mixed-precision inference. Accuracy is the meaningful contrast here.

Implemented components

• KITTI stereo image, disparity, and calibration loading
• SAD cost-volume construction along rectified scanlines
• Winner-takes-all disparity selection
• Left–right consistency checking
• Disparity-to-depth conversion using camera focal length and baseline
• Parameter analysis for block size and disparity range
• EPE, D1, depth RMSE, runtime, and valid-density evaluation
• Near/mid/far-depth and textured/textureless regional analysis
• Object-boundary, occlusion, and low-texture failure cases

Repository structure

.
├── corner_detection_harris_vs_superpoint.ipynb
├── stereo_depth_sad_vs_raft.ipynb
├── data/README.md
├── results/
│   ├── corner_detection/
│   └── stereo_depth/
├── requirements.txt
└── README.md

Setup

1. Follow data/README.md for datasets and official model dependencies.

2. Create an environment and install the notebook dependencies:

   python -m venv .venv
   source .venv/bin/activate       # Windows: .venv\Scripts\activate
   pip install -r requirements.txt
   jupyter lab

3. Start Jupyter from the repository root so the relative paths resolve correctly.

Attribution and scope

• HPatches and SuperPoint remain subject to their original licences and citations.
• KITTI data is for non-commercial research use; it is not redistributed here.
• RAFT-Stereo is an external MIT-licensed dependency and is not copied into this repository.
• Results were produced on selected benchmark training pairs and are not official benchmark-server submissions.

Author

Kübra Yurduseven