Research map / Detection and segmentation / Open-set and few-shot detection
Dual Pose-invariant Embeddings: Learning Category and Object-specific Discriminative Representations for Recognition and Retrieval
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 PiRO, a VGG-16-based dual-encoder (Pose-invariant Attention Network) that learns category and object-identity embeddings in separate spaces with pose-invariant ranking losses, improving pose-invariant recognition and retrieval on ModelNet-40, ObjectPI, and FG3D.
Problem
Pose-invariant object recognition and retrieval from single views: existing methods learn category-level embeddings in a single space and fail to separate distinct objects within the same category, so object-level tasks perform poorly.
Approach
A shared CNN backbone with two FC heads and self-attention aggregators produces category and object-identity embeddings in two separate spaces, trained jointly with pose-invariant object loss (separating confusers within a category) and pose-invariant category loss (clustering same-category objects), plus L-Softmax.
Key results
- Single-view object recognition improves by 20.0% on ModelNet40, 2.0% on ObjectPI, and 46.5% on FG3D over the previous best.
- Single-view object retrieval mAP improves by 33.7% on ModelNet-40, 18.8% on ObjectPI, and 56.9% on FG3D over the previous best.
- Dual embedding space improves average classification on ModelNet-40 by 7.7% and average retrieval by 13.0% over prior state of the art.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| learn dual category and object embeddings instead of PI-TC | PiRO · ModelNet-40 | single-view object recognition accuracy: 74.1 -> 94.1 | Helped |
| learn dual category and object embeddings instead of PI-TC | PiRO · ModelNet-40 | single-view object retrieval mAP: 41.4 -> 84.5 | Helped |
| learn dual category and object embeddings instead of PI-TC | PiRO · ObjectPI | single-view object recognition accuracy: 91.1 -> 93.1 | Helped |
| learn dual category and object embeddings instead of PI-TC | PiRO · ObjectPI | single-view object retrieval mAP: 61.8 -> 80.6 | Helped |
| learn dual category and object embeddings instead of PI-TC | PiRO · FG3D | single-view object recognition accuracy: 36.5 -> 83.0 | Helped |
| learn dual category and object embeddings instead of PI-TC | PiRO · FG3D | single-view object retrieval mAP: 15.9 -> 72.8 | Helped |
| add Lpicat to the dual-space Lcat + Lpiobj configuration (ablation) | PiRO dual embedding space · ObjectPI | single-view object recognition accuracy: 94.5 -> 92.7 | Hurt |
| add Lpicat to the dual-space Lcat + Lpiobj configuration (ablation) | PiRO dual embedding space · ModelNet-40 | single-view object retrieval mAP: 85.2 -> 84.0 | Hurt |
| add Lpicat to the dual-space Lcat + Lpiobj configuration (ablation) | PiRO dual embedding space · FG3D | single-view object recognition accuracy: 83.2 -> 83.1 | No clear effect |
Surprising findings
- Learning category and object embeddings in a single shared space creates contradicting objectives, so dual spaces are needed for gains on object-level tasks.
Limitations
- Single-view performance on ObjectPI object recognition gains are small (2.0%)
- Object-level performance depends on high embedding dimensionality (256-512 for ObjectPI)
Experimental setup
Datasets: ModelNet-40 ObjectPI FG3D
Benchmarks: ModelNet-40 ObjectPI FG3D
Metrics: accuracy mAP
Baselines: MV-CNN PI-CNN MV-TC PI-TC MV-Proxy PI-Proxy PiRO-SE
Builds on: Ho et al. pose-invariant embeddings (PI-CNN/PI-Proxy/PI-TC) L-Softmax VGG-16 proxy-based losses
Closest papers by meaning
- PIEs: Pose Invariant Embeddings (CVPR 2019 · similarity 0.93)
- NOPE: Novel Object Pose Estimation from a Single Image (CVPR 2024 · similarity 0.88)
- SecondPose: SE(3)-Consistent Dual-Stream Feature Fusion for Category-Level Pose Estimation (CVPR 2024 · similarity 0.88)
- Object Recognition as Next Token Prediction (CVPR 2024 · similarity 0.86)
- Exploiting Distribution Constraints for Scalable and Efficient Image Retrieval (ICLR 2025 · similarity 0.86)
- One2Any: One-Reference 6D Pose Estimation for Any Object (CVPR 2025 · similarity 0.86)
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.