PerSense

Training-Free Personalized Instance Segmentation in Dense Images

👑End-to-End
❄️Training-free
💡Model-agnostic
🎯One-shot Framework

Authors

Muhammad Ibraheem Siddiqui, Muhammad Umer Sheikh, Hassan Abid, Kevin Henry, Muhammad Haris Khan
PerSense segmentation demonstration

PerSense Framework

PerSense: an end-to-end, training-free and model-agnostic one-shot framework for personalized instance segmentation in dense images.

Class-label Extractor
Initial Exemplar Selection
Density Map Generation & Instance Detection
Point Prompt Selection & Initial Segmentation
Feedback Mechanism & Final Output

Class-label Extractor

Given a query image and support set, CLE leverages the masked support image to identify the object category which serves as a prompt to Grounding Detector.

PerSense Workflow

A step-by-step overview of PerSense components

Introducing PerSense++

An enhanced version of PerSense

PerSense++ Enhanced Framework Overview
Diversity-Aware Exemplar Selection
Hybrid Instance Detection Module
Irrelevant Mask Rejection Module

Diversity-Aware Exemplar Selection

Incorporates feature and scale diversity to compute a weighted score for selecting representative exemplars, enhancing density map generation

Results

Qualitative comparison with SOTA

Qualitative comparison results

Citation

If you use PerSense or PerSense++ in your research, please cite our papers

PerSense Citation
@article{siddiqui2024persense,
  title     = {PerSense: Personalized Instance Segmentation in Dense Images},
  author    = {Siddiqui, Muhammad Ibraheem and Sheikh, Muhammad Umer and Abid, Hassan and Khan, Muhammad Haris},
  journal   = {arXiv preprint arXiv:2405.13518},
  year      = {2024},
  archivePrefix = {arXiv},
  eprint    = {2405.13518},
  primaryClass = {cs.CV}
}
PerSense++ Citation
@article{siddiqui2025persense++,
    title={Towards PerSense++: Advancing Training-Free Personalized Instance Segmentation in Dense Images},
    author={Siddiqui, Muhammad Ibraheem and Sheikh, Muhammad Umer and Abid, Hassan and Henry, Kevin and Khan, Muhammad Haris},
    journal={arXiv preprint arXiv:2508.14660},
    year={2025}
  }