Research

I like problems where the thing you're looking for barely shows up in the data: a person who is a few pixels wide in a thermal image, a ship behind haze, one tree in a forest of point clouds. Most of this work happened in the PRISM research group at NIT Rourkela with Dr. Sobhan Kanti Dhara.

In plain words

Object detectors are great at finding big, clear things. They struggle when the object is tiny, low-contrast, or surrounded by clutter, and they're usually too heavy to run on a drone. My M.Tech thesis, Deep Learning Architectures for Tiny Human Detection in Aerial Thermal Imagery, tackled both problems at once: four lightweight architectures (SAPNet, TSRNet, CGANet and C2SNet), each fixing a different weak spot in the usual backbone → neck → detector pipeline.

Tiny humans in thermal imagery

Most of my M.Tech. A person seen from a drone in thermal infrared can be a blob of a few pixels. I designed lightweight detectors that still find them, cheaply enough to run on the drone itself.

  • NCC 2026 Published

    TSRNet: A Lightweight Thermal Saliency Refinement Network for Tiny Human Detection

    Vidhan Jain, Tanish, Sobhan Kanti Dhara

    A small saliency network that makes faint thermal signatures easier to see. It reaches 0.454 mAP@0.5 on RGBT TinyPerson, 5.5% above prior detectors, at only 36.5 GFLOPs. Two new modules do the heavy lifting: Dual-Frequency Refinement and Axial Context Pooling.

  • IEEE SPACE 2026 Published

    SAPNet: A Lightweight and Efficient Saliency-Aware Path Network for Tiny Human Detection in Thermal Aerial Imagery

    Vidhan Jain, Tanish, Sobhan Kanti Dhara

    Built for drones and other small, power-limited platforms: 0.404 mAP@0.5 with just 3.4M parameters and 20 GFLOPs. Saliency-path and channel-reweighting modules keep subtle heat signatures from getting lost as features are merged across scales.

  • InGaRSS 2026 Accepted

    Frequency-Aware Feature Refinement and Auxiliary Mask Supervision for Tiny Human Detection in Aerial Thermal Images

    Pratyush Sahu, Vidhan Jain, Sobhan Kanti Dhara

    Adds frequency-aware refinement and an extra mask-based training signal so the detector learns where tiny people actually are.

  • ICVGIP 2026 Under review

    Cross-Scale Correlation Guided Attention Network for Robust Tiny Human Detection in Aerial Thermal Imagery

    Vidhan Jain, Pratyush Sahu, Sobhan Kanti Dhara

    Uses correlations between feature scales to guide attention toward the few pixels that belong to a person.

  • IEEE Sensors Letters 2026 Under review

    Tiny Human Detection from Aerial Infrared Sensor Imagery via Lightweight Center-Surround Context-Aware Framework

    Vidhan Jain, Pratyush Sahu, Tanish, Sobhan Kanti Dhara

    Borrows the center-surround idea from how our own eyes pick out small objects, in a lightweight detector for infrared imagery.

Ships in messy seas

Collaborations on detecting ships in optical remote sensing images, where waves, haze and clouds make everything harder.

  • IEEE SPACE 2026 Published

    GeoStatNet: Geometric-Statistical Feature Fusion for Robust Ship Detection in Degraded Maritime Environments

    Tanish, Vidhan Jain, Sobhan Kanti Dhara

    Finding ships when the sea, haze and clouds are working against you, by fusing geometric and statistical cues.

  • IEEE GRSL 2026 Published

    GFCR-Net: Gradient-Guided and Frequency-Gated Context Refinement for Robust Ship Detection in Complex Optical Remote Sensing Images

    Tanish, Vidhan Jain, Sobhan Kanti Dhara

    Uses image gradients and frequency information to decide which surrounding context actually helps a ship detector, and which is just clutter.

Forests from LiDAR

At ISRO, I turned airborne LiDAR point clouds of dense forest into individual tree crowns using a region-growing algorithm, then checked the results against ground truth.

Recognition

  • IndiaAI Fellowship for my M.Tech research project
  • Winner, IEEE SPACE M.Tech Student Initiative Program 2026
  • Best Paper Award, IEEE SPACE (SDP-1)
  • Finalist, Best M.Tech Project Award