Research map / Label-efficient and robust learning / Recognition and long-tail
Research of Event Reconstruct Based on Multi-View Contrastive Learning (Student Abstract)
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 MCER, a multi-view supervised contrastive learning model with momentum and weight-sharing three-tower encoders for reconstructing public events from fragmented social media views, and introduces the Mul-View-data benchmark.
Problem
Event reconstruction from fragmented multi-view social media images, where information is scattered across many partial views of the same public event.
Approach
Uses supervised contrastive loss allowing multiple positives per anchor (views of the same event), with a three-tower momentum and weight-sharing encoder structure, to pull together views of the same event and push apart different events.
Key results
- MCER reaches 94.74% accuracy on NYU-data vs 91.34% for SimCLR v1 at threshold θ=0.3.
- On Mul-View-data at θ=0.3, MCER achieves 86.20% vs 85.37% for SimCLR v1.
- Ablation at θ=0.1 on NYU Depth V2: MCER 99.38% vs 93.45% without MoCo and 85.31% with unsupervised contrastive loss.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MCER instead of SimCLR v1 | NYU Depth-V2 · NYU Depth-V2, θ = 0.3 | accuracy: 91.34% -> 94.74% | Helped |
| use MCER instead of SimCLR v1 | Mul-View-data · Mul-View-data, θ = 0.3 | accuracy: 85.37% -> 86.20% | Helped |
| use MCER instead of SimCLR v1 | NYU Depth-V2 · NYU Depth-V2, θ = 0.5 | accuracy: 88.34% -> 86.42% | Hurt |
| use MCER instead of SimCLR v1 | Mul-View-data · Mul-View-data, θ = 0.5 | accuracy: 82.62% -> 82.52% | Hurt |
| add supervised contrastive loss instead of unsupervised contrastive loss (ablation) | MCER · NYU Depth-V2, θ = 0.1 | accuracy: 85.31% -> 99.38% | Helped |
| add supervised contrastive loss instead of unsupervised contrastive loss (ablation) | MCER · NYU Depth-V2, θ = 0.3 | accuracy: 82.34% -> 96.42% | Helped |
| add the MoCo momentum encoder setup (ablation) | MCER · NYU Depth-V2, θ = 0.1 | accuracy: 93.45% -> 99.38% | Helped |
| add the MoCo momentum encoder setup (ablation) | MCER · NYU Depth-V2, θ = 0.3 | accuracy: 91.74% -> 96.42% | Helped |
Surprising findings
- At threshold θ=0.5, MCER (86.42%) underperforms SimCLR v1 (88.34%) on NYU-data, and is roughly tied on Mul-View-data (82.52% vs 82.62%).
Limitations
- Scarcity of multi-view public datasets motivated a self-constructed benchmark with limited description.
- Performance is sensitive to the positive-sample threshold θ on Hamming distance, and gains vanish at some settings.
Experimental setup
Datasets: NYU Depth V2 Mul-View-data
Benchmarks: NYU Depth V2 Mul-View-data
Metrics: accuracy
Baselines: SimCLR v1 Remove MoCo variant (unsupervised contrastive loss)
Builds on: SimCLR MoCo
Closest papers by meaning
- CLIP-Event: Connecting Text and Images With Event Structures (CVPR 2022 · similarity 0.84)
- Self-Weighted Contrastive Learning among Multiple Views for Mitigating Representation Degeneration (NeurIPS 2023 · similarity 0.83)
- REBAR: Retrieval-Based Reconstruction for Time-series Contrastive Learning (ICLR 2024 · similarity 0.83)
- EventDance: Unsupervised Source-free Cross-modal Adaptation for Event-based Object Recognition (CVPR 2024 · similarity 0.82)
- VidEvent: A Large Dataset for Understanding Dynamic Evolution of Events in Videos (AAAI 2025 · similarity 0.81)
- Multi-View Causal Representation Learning with Partial Observability (ICLR 2024 · similarity 0.81)
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.