Research

IDTCount: Counting Every Instance Once in Open Vocabulary Video Counting

Joana Konadu Owusu, Shivanand Venkanna Sheshappanavar

WACV 2027 · Under review

Integrates a learned re-identification module and an association mechanism into SAM3’s video tracking loop. A projection head trained with supervised contrastive learning learns compact identity representations to help count each instance once.

SAM3Count for Zero-Shot Open Vocabulary Counting in Images and Videos

Joana Konadu Owusu, Shivanand Venkanna Sheshappanavar

WiCV · CVPR Workshops 2026 · Oral

A text-prompted counting framework built on SAM3. Density-aware processing applies region-guided adaptive tiling to crowded images, while a multimodal re-identification tracker combines appearance, motion, spatial, and temporal cues to preserve object identities in video.

Object Counting Across Modalities: Taxonomies, Benchmarks, Applications, and Open Challenges

Joana Konadu Owusu, Shivanand Venkanna Sheshappanavar

TPAMI 2027 · Under review · arXiv preprint

Surveys object-counting methods through a five-axis taxonomy covering modality, mechanism, prompting, supervision, and generalization. It examines applications and benchmark limitations, then outlines directions for stronger multimodal evaluation and reliable counting beyond familiar datasets.

XPT: Enhanced Point Transformer V3

Joana Owusu

Project report · May 2025

Builds on Point Transformer V3 with Dynamic Tanh, QKV normalization, and 3D rotary positional embeddings. These changes target faster, more stable point-cloud understanding, evaluated on ScanNet v2 and S3DIS for segmentation accuracy and computational efficiency.