Implementation of classical and modern computer vision algorithms developed during the Computer Vision course at FAU Erlangen-Nürnberg (Summer 2025).
- RANSAC: Plane detection in 3D point clouds
- PRANSAC (Preemptive RANSAC): Early termination for efficiency
- MLESAC: Maximum likelihood estimation variant
- Application: Box detection from ToF depth data
- VLAD Encoding: Vector of Locally Aggregated Descriptors
- Custom SIFT: Hellinger-normalized descriptors with angle forcing
- Generalized Max Pooling: Ridge regression-based aggregation
- Multi-VLAD + PCA: Multiple codebooks with whitening
- E-SVM: Exemplar SVM for refinement
- Dataset: ICDAR2017 Historical Writer Identification
- Region Proposals: Hierarchical segmentation merging
- Similarity Metrics: Color, texture, size, fill
- Object Detection: Region-based detection pipeline
- Dataset: Archaeological artifacts (Art History, Christian & Classical Archaeology)
- Bayer Pattern Demosaicing: RGB reconstruction from RAW sensor data
- HDR Fusion: Combining multiple exposures
- Tone Mapping: iCAM06 algorithm
- White Balance: Gray world implementation
- Face Detection: MTCNN implementation
- Face Tracking: Template matching for video
- Face Recognition: FaceNet embeddings with k-NN
- Face Clustering: k-means for person re-identification
- Evaluation: DIR curves for open-set identification
Python NumPy OpenCV scikit-learn SciPy Matplotlib MTCNN FaceNet rawpy
- Exercise 1: Robust plane detection using aforementioned algorithms
- Exercise 2: 0.75 mAP on Writer Identification, 0.88 Top-1 Accuracy on Writer Retrieval
- Exercise 3: ~2000 region proposals with high recall
- Exercise 4: Natural HDR images merged from RAW data with proper color reproduction
- Exercise 5: >90% accuracy on closed-set face identification