Research map / Detection and segmentation / CNN features
GigaHumanDet: Exploring Full-Body Detection on Gigapixel-Level Images
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GigaHumanDet, a corner-based full-body detector for gigapixel images that matches top-left and bottom-right corners using instance-guided contrastive embeddings and a shape-aware bodyness score with multi-precision width/height regression.
Problem
Full-body person detection in gigapixel-level images, where center-regression detectors struggle due to diverse postures (unstable centers) and large scale variation (hard long-range width/height regression).
Approach
Replace center-based box encoding with corner pairs (each needing only two boundaries), and match corners using a shape-aware bodyness score built from body width/height embeddings regressed at corner locations via a Gaussian-inspired multi-precision strategy, plus an instance-guided contrastive loss to make corner embeddings discriminative.
Key results
- GigaHumanDet with Hourglass-104 reaches 82.4% AP50 on PANDA (single-scale), vs 21.6% for CornerNet baseline and 68.0% for DINO.
- On STCrowd, GigaHumanDet achieves 90.7% APS T50 with Hourglass-104, surpassing CornerNet baseline by 22.3 points.
- Gaussian-inspired multi-precision regression beats uniform-inspired (79.1 vs 78.8 AP50) and single regression (78.1) under Hourglass-52; shape-aware term adds 0.8 AP50.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add instance-guided learning (ablation) | GigaHumanDet · PANDA | AP50: 79.1 -> 79.6 | Helped |
| replace Gaussian-inspired multi-precision regression with uniform-inspired multi-precision regression (ablation) | GigaHumanDet · PANDA | AP50: 79.1 -> 78.8 | Hurt |
| replace multi-precision regression with single regression (ablation) | GigaHumanDet · PANDA | AP50: 79.1 -> 78.1 | Hurt |
| add the shape-aware term to bodyness (ablation) | GigaHumanDet · PANDA | AP50: 78.3 -> 79.1 | Helped |
| set bodyness threshold to 0.45 instead of 0.4 (ablation) | GigaHumanDet · PANDA | AP50: 82.1 -> 82.4 | Helped |
| vary the number of retained corner keypoints (ablation) | GigaHumanDet · PANDA | AP50: 80.6 -> 82.4 -> 81.6 -> 81.4 -> 80.8 | Mixed |
Surprising findings
- Naive CornerNet baseline achieves only 21.6% AP50 on PANDA, showing heuristic corner matching fails at gigapixel scale.
Limitations
- Accuracy depends on keeping top-k keypoints (k=200 chosen), a heuristic that the authors note is sensitive to crowd density inhomogeneity.
- Bodyness threshold (0.45) and top-k are tuned hyperparameters on the evaluation set.
Experimental setup
Datasets: PANDA STCrowd
Benchmarks: PANDA STCrowd
Metrics: AP50 AP75 APS50 APM50 APL50 AR #Pass
Baselines: Faster R-CNN Cascade R-CNN RetinaNet ATSS YOLOX DINO ClusDet PAN CSP Pedestron CenterNet CentripetalNet
Builds on: CornerNet CenterNet corner pooling Focal Loss
Closest papers by meaning
- Detangling People: Individuating Multiple Close People and Their Body Parts via Region Assembly (CVPR 2017 · similarity 0.87)
- GigaMoE: Sparsity-Guided Mixture of Experts for Efficient Gigapixel Object Detection (AAAI 2026 · similarity 0.86)
- DeepCut: Joint Subset Partition and Labeling for Multi Person Pose Estimation (CVPR 2016 · similarity 0.86)
- PANDA: A Gigapixel-Level Human-Centric Video Dataset (CVPR 2020 · similarity 0.85)
- PairDETR : Joint Detection and Association of Human Bodies and Faces (CVPR 2024 · similarity 0.85)
- PBADet: A One-Stage Anchor-Free Approach for Part-Body Association (ICLR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.