Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Co-Domain Embedding Using Deep Quadruplet Networks for Unseen Traffic Sign Recognition
Junsik Kim, Seokju Lee, Tae-Hyun Oh, In So Kweon
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 |
|---|---|---|---|
| Use template-anchored quadruplet training with HingeM-5 instead of Triplet-DA | Traffic-sign one-shot recognition · GTSRB-sub unseen classes | Top 1-NN accuracy (%): 47.6 -> 47.8 | No clear effect |
| Use template-anchored quadruplet training with HingeM-5 instead of Triplet-DA | Traffic-sign one-shot recognition · TT100K-c unseen classes | Top 1-NN accuracy (%): 67.1 -> 67.3 | No clear effect |
| Use template-anchored quadruplet training with HingeM-5 instead of IdsiaNet | Traffic-sign one-shot recognition · GTSRB unseen classes | Top 1-NN accuracy (%): 40.2 -> 45.2 | Improved |
| Train quadruplet embeddings with HingeM instead of IdsiaNet features | Cross-dataset traffic-sign transfer · TT100K-c sample average (trained on GTSRB) | Top1 NN accuracy (%): 36.5 -> 42.3 | Improved |
| Use 3 pairwise loss terms instead of HingeM-5 (ablation) | Quadruplet network · GTSRB-all validation set | Top1 NN accuracy, Avg. (%): 69.1 -> 67.3 | Hurt |
| Add real-real pairwise loss term to HingeM-5 (ablation) | Quadruplet network · GTSRB-all validation set | Top1 NN accuracy, Unseen (%): 41.6 -> 39.2 | Hurt |
| Add real-real pairwise loss term to HingeM-5 (ablation) | Quadruplet network · GTSRB-all validation set | Top1 NN accuracy, Seen (%): 95.3 -> 97.3 | Improved |
| Set embedding dimension to 50 instead of 100 (ablation) | Quadruplet network with HingeM-5 · GTSRB-all validation set | Top1 NN accuracy, Avg. (%): 69.1 -> 67.1 | Hurt |
| Set embedding dimension to 150 instead of 100 (ablation) | Quadruplet network with HingeM-5 · GTSRB-all validation set | Top1 NN accuracy, Avg. (%): 69.1 -> 68.9 | Hurt |
| Replace HingeM pairwise losses with contrastive-style losses (ablation) | Quadruplet network · GTSRB-all unseen-class representation, 200 SVM instances/class | SVM classification error (%): 3.77 ($\pm$0.29) -> 5.30 ($\pm$0.14) | Hurt |
| Replace HingeM pairwise losses with contrastive-style losses (ablation) | Quadruplet network · GTSRB-sub unseen-class representation, 200 SVM instances/class | SVM classification error (%): 13.81 ($\pm$0.31) -> 12.12 ($\pm$0.33) | Improved |
Closest papers by meaning
- Traffic-Sign Detection and Classification in the Wild (CVPR 2016 · similarity 0.84)
- The Mapillary Traffic Sign Dataset for Detection and Classification on a Global Scale (ECCV 2020 · similarity 0.81)
- Multi-Object Tracking With Quadruplet Convolutional Neural Networks (CVPR 2017 · similarity 0.80)
- DI-V2X: Learning Domain-Invariant Representation for Vehicle-Infrastructure Collaborative 3D Object Detection (AAAI 2024 · similarity 0.80)
- X3D: Expanding Architectures for Efficient Video Recognition (CVPR 2020 · similarity 0.80)
- Yet Another Traffic Classifier: A Masked Autoencoder Based Traffic Transformer with Multi-Level Flow Representation (AAAI 2023 · similarity 0.80)
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.
