Muqsit Islam

A sliding-window SLAM pipeline built on VGGT’s feed-forward geometry. Local submaps are aligned by camera pose, so the map stays consistent over long sequences while GPU memory stays under 17 GB. 2D instance masks are lifted to persistent 3D objects for change detection and placed on an estimated floor plane for navigation.

I implemented pipeline components, ran the ScanNet++ benchmarks, and worked on the analysis and revision.

I compared seven harmonization methods (GAN, diffusion and statistical) on 1,112 brain MRI scans from 17 sites. The usual score, a frozen site classifier, turned out to be unreliable: classifiers trained with the benchmark’s own recipe disagreed by up to 0.46 balanced accuracy on the same outputs, retrained classifiers still found the site, and the best score went to the most destructive method.

With Dr. Pierrick Coupé (CNRS, LaBRI) and Dr. José M. Martínez (UAM).

A ViT-L re-identification system for the ICIP 2026 challenge track. I tried semantic supervision, class-aware retrieval, DINOv2/CLIP reranking and feature fusion, but the biggest gains came from simpler choices: keeping each image’s aspect ratio and training robust embeddings, which improved leaderboard mAP by about 5% relative.

Two WWF-supported systems from my time at the LUMS Computer Vision & Graphics Lab: species detection from camera traps, and early fire detection from remote camera feeds. I fine-tuned MegaDetector, YOLO-family models and CLIP on field data to cut false alarms from environmental motion, and built the inference, visualization and monitoring around them so they could run unattended.