ICPR 2026 Competition on Privacy-Preserving Person Re-Identification from Top-View RGB-Depth Camera (TVRID)

Technical Summary: ICPR 2026 TVRID Competition

Problem Statement
Person Re-Identification (Re-ID) traditionally relies on RGB imagery, where identity is inferred from clothing color and texture. However, this approach faces two critical limitations: vulnerability to real-world factors like viewpoint changes and occlusion, and significant privacy concerns in public surveillance due to the exposure of personally identifiable appearance cues. While depth sensing offers a privacy-preserving alternative by capturing 3D structure and suppressing color-based cues, existing top-view datasets lack support for controlled geometric variations (e.g., ascent and descent) that significantly alter perceived body proportions. The ICPR 2026 TVRID (Top-View RGB-Depth Person Re-Identification) competition addresses this gap by evaluating Re-ID performance under privacy-aware, modality-constrained, and cross-modal conditions.

Dataset and Competition Setup
The competition is built upon the TVRID dataset, comprising 86 identities captured by four synchronized overhead Intel RealSense D455 cameras. The acquisition setup includes structured geometric variations across four distinct viewpoints:

1. Flat: Canonical level-ground passage.
2. Ascent: Upward motion on a stepladder.
3. Descent: Downward motion on a stepladder.
4. Oblique: Elevated oblique roof view.

Each subject is recorded twice per camera (IN/OUT directions), resulting in paired, synchronized RGB and Depth streams at 15 FPS. The dataset includes pre-computed person crops to facilitate end-to-end pipelines.

The competition evaluated three distinct tracks:

• Track 1 (RGB Re-ID): Standard Re-ID using RGB images, serving as an upper-bound baseline.
• Track 2 (Depth Re-ID): Privacy-preserving Re-ID using depth maps only, testing robustness against geometric shifts without color cues.
• Track 3 (RGB↔Depth Cross-Modal Re-ID): Retrieval where the query and gallery belong to different modalities (e.g., RGB query vs. Depth gallery), testing modality-invariant representation learning.

Evaluation was conducted using CMC@1 (Cumulative Matching Characteristic at Rank-1) and mAP (mean Average Precision) across three camera-based scenarios: within-camera (IN/OUT), ascent vs. descent, and flat vs. oblique.

Methodological Approaches of Top Entries
The competition attracted 147 participants, with 33 final submissions on the hidden test set. The top-performing methods across tracks utilized advanced deep learning architectures and specific training strategies:

• RGB Track:

  • Hien Pham Duy et al. achieved 100% CMC@1 using a ViT-Base backbone with identity-aware PK sampling and human-centered augmentations (e.g., background erasing, body-part erasing). They employed mean pooling for sequence aggregation and k-reciprocal re-ranking.
  • Carlo Metta utilized a ConvNeXt-Base backbone with ArcFace and batch-hard triplet losses, employing 5-fold cross-validation and ensembling.
  • Oron Nir et al. applied MINER (Modeling Inter- and Intra-Person Variance), a contrastive learning framework built on EVA-CLIP DIFFER, focusing on hard negative/positive mining to learn view-invariant representations.

• Depth Track:

  • Jin-Hui Jiang et al. (98.86% CMC@1) adapted VSLA-CLIP with a ViT-Base/16 backbone. They used a two-stage cyclic optimization involving trainable temporal adaptation modules and cross-modal losses (supervised contrastive, MSE distillation, InfoNCE) to indirectly shape depth embeddings via RGB supervision.
  • Oron Nir et al. adapted the MINER framework by converting depth maps to inverse depth, normalizing them, and applying a Spectral colormap to create pseudo-RGB inputs compatible with the vision transformer backbone.
  • Md Rashidunnabi et al. proposed a sequence-based pipeline using ResNet-50 with temporal attention pooling, trained with a combination of identity classification, triplet losses, and center loss on paired RGB-Depth data.

• Cross-Modal Track:

  • Md Rashidunnabi et al. (99.84% CMC@1) employed XM-VSLA, an adaptation of VSLA-CLIP that treats RGB-Depth Re-ID as a vision-language alignment problem within a shared embedding space, guided by identity-conditioned text prompts.
  • Jin-Hui Jiang et al. and Oron Nir et al. similarly leveraged VSLA-CLIP and MINER frameworks, respectively, utilizing explicit cross-modal contrastive alignment and dual-encoder architectures to minimize the discrepancy between RGB and depth representations.

Key Results
The final results established a clear difficulty hierarchy: RGB > Depth > Cross-Modal.

• RGB Track: Top systems achieved near-perfect performance (100% CMC@1, 100% mAP).
• Depth Track: Performance dropped significantly compared to RGB, with the top system achieving 98.86% CMC@1 and 99.38% mAP. The aggregate median mAP for the track was 49.82%, indicating a substantial challenge for depth-only systems.
• Cross-Modal Track: This was the most difficult task. The top system achieved 99.84% CMC@1 and 99.92% mAP, but the aggregate median mAP was only 23.30%. The performance gap between Depth and Cross-Modal was approximately 16.44 points in mAP.

Qualitative analysis suggested that matching flat-view samples against the oblique camera was the most difficult configuration.

Significance and Claims
The paper positions TVRID as a reproducible benchmark for top-view, depth-based, and cross-modal person Re-ID. Its primary contributions are:

1. Dataset: Providing a synchronized RGB-Depth dataset with structured geometric variations (flat, ascent, descent, oblique) that are common in real-world deployments but underrepresented in prior benchmarks.
2. Protocol: Establishing a unified evaluation framework that simultaneously assesses recognition performance, modality robustness, and privacy-performance trade-offs.
3. Insights:
   • The results confirm that while depth sensing reduces average retrieval accuracy compared to RGB, strong modality-invariant representations are achievable.
   • The success of cross-modal retrieval (RGB query $\to$ Depth gallery) challenges the assumption that depth is inherently privacy-preserving. The authors argue that if depth can be reliably linked to RGB identity signatures, privacy protection becomes a conditional property dependent on threat models and access control, rather than an automatic guarantee.
   • Camera geometry (specifically elevation changes) remains a key stress factor for Re-ID systems.

The paper concludes that TVRID serves not only as an evaluation benchmark but also as a framework for discussing operational privacy risks in modern Re-ID systems, highlighting the need for nuanced interpretations of privacy in the context of cross-modal linkage.