Projects
GitHub repositories, work and academic projects. A few run in your browser.
| Project | In short | Type | Tags | |
|---|---|---|---|---|
| mAP Calculator | Mean average precision by PascalVOC 2012 and COCO rules. | proj | Evaluation and metricsObject detectionDeveloper tools | Run in browser → |
| FreeMatch in PyTorch | Unofficial implementation of the FreeMatch algorithm for semi-supervised image classification. | proj | Semi-supervised learningImage classification | GitHub ↗ |
| Cross-Domain Adaptive Clustering | Contributed the Cross-Domain Adaptive Clustering algorithm to Dassl.pytorch for domain-adaptive image classification. | proj | Domain adaptationSemi-supervised learningImage classification | GitHub ↗ |
| Multi-Radius Deep SVDD | Multi-radius Deep SVDD, an improvement on deep one-class anomaly detection. | proj | Anomaly detection | GitHub ↗ |
| CTAugment for PyTorch | CTAugment wrapper for PyTorch classification models. | proj | Semi-supervised learningData augmentation | GitHub ↗ |
| Vehicle Counting UI | Interface for vehicle counting and camera calibration. | proj | Traffic analyticsDeveloper tools | GitHub ↗ |
| Intelligent Traffic Light Control System | Prototype that automates traffic-light timers at crossroads from vehicle density, measured with YOLO detection and with foreground extraction. | proj | Traffic analyticsObject detection | GitHub ↗ |
| Thread Management System | Thread management library that creates, schedules and kills threads at kernel level with Linux system calls, similar to pthread. | proj | Systems programming | GitHub ↗ |
| Saliency Map Generation from Images | Saliency maps by a supervised approach (superpixel features with random forest regression) and a semi-supervised autoencoder with attention. | proj | Saliency | GitHub ↗ |
| Classification of Gender from Tweets | Gender classification from tweets and profile descriptions, using NLP for formatting, C4.5 decision trees for features and shallow neural networks for classification. 84% accuracy on the test set. | proj | Natural language | — |
| Redefining the One-Shot Object Detector for a Two-Class Problem | AP head 88.5% → 91%, AP body 87.5% → 91.9% | proj | Object detectionTraining recipes | Read → |
| Low-Compute, High-FPS Background Subtraction | 62 FPS with one instance on an Intel Celeron N3350 | proj | Edge deploymentTraffic analytics | Read → |
| Site-Specific Training Tool | mAP 64.9 → 74.1 with 250–300 labeled images | proj | Object detectionDomain adaptation | Read → |
| Vehicle Trajectory Anomaly Detection | Anomalous trajectories: precision 0.95, recall 1.00 | proj | Anomaly detectionTrackingTraffic analytics | Read → |
| Person Retrieval and Re-Identification | Market1501 mAP 85.24 → 91.24 | proj | Re-identificationTracking | Read → |
| Unsupervised Domain Adaptation and Semi-Supervised Learning | About 6% accuracy gain on long-tailed private data | proj | Domain adaptationSemi-supervised learningClass imbalance | Read → |
| Network Pruning of Object Detectors | 2–3× speed-up, 3–4× less memory on embedded devices | proj | PruningEdge deploymentObject detection | Read → |
| INT8 Quantization of Object Detectors | 0.5% drop after INT8 quantization | proj | QuantizationEdge deploymentObject detection | Read → |
18 projects.