S M Rayeed
Home

Research

Six lines of work, from species identification and biological VQA to knowledge-grounded retrieval. Each area has a page with every project, its summary and an overview figure.

Curated collectionsBeetlePaloozaBeetlePUUMNHM-CarabidsTransfer holds between lab collectionsIn the fieldI1MC: varied pose, occlusion, cluttertrain37%performance drop fromlab-trained modelsViLT + MLP: 97% genus, 94% species accuracy on merged data.50% of the training set gives nearly the same accuracy.

Fine-grained taxonomic identification

Which vision models can tell 1,769 beetle species apart, how little data they need, and where they break.

3 papersExplore this area
AuditRepairAugmentQ: When was it built?A: 1889Knowledge base: no dateQuestion: which building?Answers unsupported by theknowledge base:22%of InfoSeekUnderspecified questions:59%E-VQA47%InfoSeekQ: When was the iron latticetower in this photo built?A: 1887 to 1889Evidence: KB passage 3Derivable answers, clearquestions. Human accuracy:92.9E-VQA91.5InfoSeekdistractor entityA second entity in the image.InfoSeek retrieval R@1:43.514.7QA accuracy of strong systemsroughly halves (IBA 40.1 to 21.4)Model rankings change once benchmarks are repaired and augmented.

Knowledge-based VQA and retrieval

Benchmarks that measure real knowledge-grounded reasoning, and retrieval systems that know when evidence is not enough.

3 papers, work in progressExplore this area
Tray photographIterative detectionGrounding DINO + SAM; found boxes aremasked and the tray is re-detected atlower thresholds, 0.30 down to 0.10VLM check:any left?yes: flag for recheckCropped and ordered123456Scale bar and color cardlocated with MoondreamNEON recordsMetadata validatedagainst the databaseCuration errors returned to the biorepository65,236specimens3,649trays518species96%tray-level detection

AI-ready biodiversity data

Turning museum drawers and NEON traps into validated, machine-learning-ready datasets.

4 papers, 2 datasetsExplore this area
21 keypoints, x, y and depth63featuresANN, 4 hidden layersঅআইউএওকখগঘচছজঝটঠ০১২৩৪৫৬৭47 classes47,000images10signers37 + 10letters and digits97.84%F1 score

Sign language recognition

Depth-based recognition of Bangla Sign Language, and the first complete open dataset for it.

2 papers, 1 datasetExplore this area