Skip to content
DSML.KZ
Publications

Code · Zenodo · August 25, 2026

DSTA: Density-Steered Token Attention — code, benchmarks, mechanistic probes and complete run logs

Arulmurugan Ramu, Altynzer Baiganova, Sowjanya Lakshmi A.

Arulmurugan Ramu
Open via DOI

Abstract

Reference implementation, benchmark construction, mechanistic probes (P1-P6) and the complete raw run logs behind every table and figure in the article "Mechanistic Evaluation of Transformer-based Architectures for Enhanced Small Object Detection in Crowded Computer Vision Scenes: Density-Steered Token Attention" (Computer Vision and Image Understanding). Includes: the DSTA encoder and every baseline variant; the CrowdSynth-D generator; the COCO-SOC selection criterion, image identifiers and construction script; 35 per-run JSON records with metrics, probe measurements and full loss histories; 41 training logs; and the analysis code that regenerates every table and figure from those logs. This deposit contains the code and the data only; the manuscript is not included. Code is released under MIT; run logs and the tables and figures generated from them under CC BY 4.0. MS-COCO images are not redistributed.

Advanced Neural Network ApplicationsVisual Attention and Saliency DetectionDomain Adaptation and Few-Shot Learningsmall object detectioncrowded scenestransformer models

More by Arulmurugan Ramu