Research map / Label-efficient and robust learning / Recognition and long-tail
Negative Samples Are at Large: Leveraging Hard-Distance Elastic Loss for Re-identification
Hyungtae Lee, Sungmin Eum, Heesung Kwon
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add HE loss instead of Triplet loss (ablation) | re-ID baseline · VeRi-776 | mAP: 79.7 -> 80.6 | Improved |
| add MoReID dictionary with Triplet loss (ablation) | re-ID baseline · VeRi-776 | mAP: 79.7 -> 79.4 | Hurt |
| combine MoReID with HE loss (ablation) | re-ID baseline · VeRi-776 | mAP: 79.7 -> 83.3 | Improved |
| increase dictionary size from 1024 to 8192 with HE loss (ablation) | MoReID · VeRi-776 | mAP: 81.9 -> 83.3 | Improved |
| increase dictionary size with Triplet losses (ablation) | MoReID Tri-all · VeRi-776 | mAP: 79.7 -> 79.3 | No clear effect |
| raise momentum m from 0.9 to 0.997 (ablation) | MoReID · VeRi-776 | mAP: 79.5 -> 83.3 | Improved |
| use batch configuration 16x16 instead of 128x2 (ablation) | HE loss · VeRi-776 | mAP: 77.9 -> 83.3 | Improved |
| replace HE loss with N-pair loss (ablation) | MoReID · VeRi-776 | mAP: 83.3 -> 79.8 | Hurt |
| replace HE loss with Ranked List loss (ablation) | MoReID · VeRi-776 | mAP: 83.3 -> 80.5 | Hurt |
| replace HE loss with vanilla InfoNCE (ablation) | MoReID · VeRi-776 (16x16) | mAP: 83.3 -> 70.5 | Hurt |
| add hard mining to InfoNCE (ablation) | InfoNCE · VeRi-776 (128x2) | mAP: 62.9 -> 63.4 | Improved |
| add past positive samples to the dictionary (ablation) | MoReID + HE loss · VeRi-776 | mAP: 83.3 -> 80.7 | Hurt |
Closest papers by meaning
- Adaptive Sparse Pairwise Loss for Object Re-Identification (CVPR 2023 · similarity 0.86)
- Large-Scale Training Data Search for Object Re-Identification (CVPR 2023 · similarity 0.85)
- Difficulty-Based Sampling for Debiased Contrastive Representation Learning (CVPR 2023 · similarity 0.83)
- Alice Benchmarks: Connecting Real World Re-Identification with the Synthetic (ICLR 2024 · similarity 0.83)
- Contrastive Learning with Hard Negative Samples (ICLR 2021 · similarity 0.83)
- The Dilemma of TriHard Loss and an Element-Weighted TriHard Loss for Person Re-Identification (NeurIPS 2020 · similarity 0.83)
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.
