Research map / Detection and segmentation / 3D detection and driving
Disentangling 3D Prototypical Networks for Few-Shot Concept Learning
Mihir Prabhudesai, Shamit Lal, Darshan Patil, Hsiao-Yu Tung, Adam W Harley, Katerina Fragkiadaki
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 disentangled 3D shape and style prototypes instead of the 3DP-Net representation without object disentanglement | few-shot concept classification · CLEVR, Real Veggie, and Replica | accuracy | Improved |
| use D3DP-VQA rather than a 2D NSCL model | CLEVR VQA · one-shot novel-object test, 10 training examples | accuracy: 0.594 -> 0.775 | Improved |
| use 2D NSCL rather than D3DP-VQA | CLEVR VQA · in-domain test, 250 training examples | accuracy: 0.990 -> 0.939 | Hurt |
| add 3D shape prototypes instead of spatially pooled 1D shape codes (ablation) | D3DP-VQA · CLEVR VQA, one-shot novel-object test, 10 training examples | accuracy: 0.410 -> 0.775 | Improved |
| add shape/style disentanglement (ablation) | D3DP-VQA · CLEVR VQA, one-shot novel-object test, 10 training examples | accuracy: 0.457 -> 0.775 | Improved |
| add both 3D shape prototypes and shape/style disentanglement (ablation) | D3DP-VQA · CLEVR VQA, one-shot novel-object test, 10 training examples | accuracy: 0.608 -> 0.775 | Improved |
| replace disentangled features with entangled features (ablation) | CLEVR VQA · one-shot novel-object test, 10 training examples | accuracy: 0.619 -> 0.775 | Hurt |
| use InstanceNorm disentangled features with a rotation-aware check rather than the full method (ablation) | CLEVR VQA · one-shot novel-object test, 250 training examples | accuracy: 0.845 -> 0.861 | Improved |
| add generated imagined scenes to real detector training data | 3D object detector · CARLA, 50 annotations, IoU .3 | mean average precision: 0.57 -> 0.61 | Improved |
| add generated imagined scenes to real detector training data | 3D object detector · CLEVR, 50 annotations, IoU .3 | mean average precision: 0.43 -> 0.44 | Improved |
| increase imagined object count and randomize object locations, rather than relying on unique shape-style compositions | 3D object detector augmentation · CARLA and CLEVR | mean average precision | Improved |
Closest papers by meaning
- Few-Shot Single-View 3-D Object Reconstruction with Compositional Priors (ECCV 2020 · similarity 0.89)
- CoCoNets: Continuous Contrastive 3D Scene Representations (CVPR 2021 · similarity 0.88)
- Learning Unsupervised Hierarchical Part Decomposition of 3D Objects From a Single RGB Image (CVPR 2020 · similarity 0.88)
- Common3D: Self-Supervised Learning of 3D Morphable Models for Common Objects in Neural Feature Space (CVPR 2025 · similarity 0.87)
- Object-Centric Representation Learning for Enhanced 3D Semantic Scene Graph Prediction (NeurIPS 2025 · similarity 0.87)
- ImageNet3D: Towards General-Purpose Object-Level 3D Understanding (NeurIPS 2024 · 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.
