Wildlife Grid Detection using Hand-Crafted Features (Course Project)
OtherImageClassification
The wildlife grid detection project involved labeling animal presence across grid cells in wildlife images. I implemented confidence-based pseudo-labeling to expand the training set from limited labeled data and refined detection accuracy via label calibration. The primary goal was to classify each grid cell in 800×600 images for animal presence using a semi-supervised approach.
• Applied hand-crafted feature engineering (HSV, Canny edge, HOG, LBG) to improve labeling precision.
• Used semi-supervised learning and XGBoost for confidence-based pseudo-labeling.
• Enhanced spatial label consistency with probability calibration and post-processing.
• Achieved 82.7% accuracy through label refinement processes.
The wildlife grid detection project involved labeling animal presence across grid cells in wildlife images. I implemented confidence-based pseudo-labeling to expand the training set from limited labeled data and refined detection accuracy via label calibration. The primary goal was to classify each grid cell in 800×600 images for animal presence using a semi-supervised approach. • Applied hand-crafted feature engineering (HSV, Canny edge, HOG, LBG) to improve labeling precision. • Used semi-supervised learning and XGBoost for confidence-based pseudo-labeling. • Enhanced spatial label consistency with probability calibration and post-processing. • Achieved 82.7% accuracy through label refinement processes.
2025 - 2025