Hengyi traditional classifier3

[HollyXu26]

Fix:

• traditional_classifier.ipynb's comparison table was using the nearest-centroid baseline's validation-set accuracy (2.84%/8.24%) instead of test-set accuracy, making the SVM/RF-vs-baseline comparison inconsistent. Added a dedicated test-set baseline (2.24%/7.26%) and corrected the table and Summary accordingly.

New:

• Added sift_bovw_classifier.ipynb: a third handcrafted descriptor (SIFT keypoints encoded as Bag-of-Visual-Words), self-contained pipeline covering:
  • MiniBatchKMeans visual vocabulary (300 words) + BoVW histogram encoding, cached to outputs/*_sift_bovw_features.npz
  • Nearest-centroid baseline, Linear SVM (C grid search on validation), and Random Forest, evaluated once on the test set
  • Confusion matrix, hardest-class breakdown, most-confused species pairs, and misclassified-image visualisation for both classifiers
  • Optional section concatenating SIFT-BoVW with the existing HOG + colour feature and retraining a Linear SVM

Key findings:

• SIFT-BoVW's baseline (3.58%/11.80%) and Linear SVM (3.72%/10.24%, best top-1 of any traditional config so far) both beat their HOG + colour counterparts.
• Unlike HOG + colour, Random Forest underperforms its own baseline on SIFT-BoVW (2.64% vs 3.58%) — the same hyperparameters don't transfer across feature spaces; discussed as future work rather than retuned here.
• Concatenating HOG + colour + SIFT-BoVW (2.96% top-1) does not beat SIFT-BoVW alone, likely because the shared PCA is dominated by the much larger HOG + colour block.

Docs:

• classifier/README.md: added SIFT-BoVW results/discussion, restructured so each notebook's results live in its own subsection (previously a separate "## Results" heading only covered HOG + colour).
• traditional/README.md: updated the top-level index to mention SIFT feature extraction and the three feature representations now compared.