Research map / Vision-language and multimodal / Human and embodied modeling
LucidPPN: Unambiguous Prototypical Parts Network for User-centric Interpretable Computer Vision
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 LucidPPN, a prototypical parts network with separate grayscale shape/texture and color branches, so explanations disentangle color from shape and texture, with accuracy comparable to PIP-Net and a user study favoring it.
Problem
Prototypical-parts networks are ambiguous because each image patch mixes color, shape, and texture features, so users cannot tell which feature drives a decision.
Approach
Two branches: ShapeTexNet operates on grayscale images to capture shape and texture, while ColorNet uses a downscaled image with 1x1 convolutions per pixel to capture color only; their resemblance maps are multiplied per object part and class, and prototypes are aligned to object parts using PDiscoNet masks.
Key results
- LucidPPN reaches 91.6±0.2% on Stanford Cars and 95.0±0.3% on Oxford Flowers, the best among compared PP models, while 81.5±0.4% on CUB is below PIP-Net's 84.3±0.2%.
- Under hue perturbation on CUB, PIP-Net accuracy drops from 83.9 to 53.0, LucidPPN from 81.9 to 71.7, and ShapeTexNet stays at 80.3.
- In a 60-worker user study, users guided by LucidPPN explanations reached 67.9±16.9% accuracy versus 60.0±18.1% with PIP-Net (p=0.044).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fuse ColorNet with ShapeTexNet (ablation) | LucidPPN · CUB | top-1 accuracy: 80.4 -> 81.8 | Helped |
| Fuse ColorNet with ShapeTexNet (ablation) | LucidPPN · CARS | top-1 accuracy: 91.7 -> 91.7 | No clear effect |
| Fuse ColorNet with ShapeTexNet (ablation) | LucidPPN · DOGS | top-1 accuracy: 78.6 -> 78.9 | Helped |
| Fuse ColorNet with ShapeTexNet (ablation) | LucidPPN · FLOWER | top-1 accuracy: 93.6 -> 95.3 | Helped |
| Use LucidPPN rather than PIP-Net | CUB classifiers under hue perturbation · CUB hue-perturbed | top-1 accuracy: 53.0 -> 71.7 | Helped |
| Use ShapeTexNet without color input (ablation) | CUB classifier under hue perturbation · CUB hue-perturbed | top-1 accuracy: 80.3 -> 80.3 | No clear effect |
| Use LucidPPN explanations instead of PIP-Net explanations | CUB user study participants · CUB user study | mean user accuracy: 60.0±18.1 -> 67.9±16.9 | Helped |
| Increase the number of object parts K (ablation) | LucidPPN · CUB, CARS, DOGS, and FLOWER | top-1 accuracy | Helped |
| Increase the correspondence-loss weight α_D (ablation) | LucidPPN · evaluated datasets | intersection-over-union | Helped |
| Increase the correspondence-loss weight α_D beyond its accuracy-optimal range (ablation) | LucidPPN · evaluated datasets | top-1 accuracy | Hurt |
| Include the prototypical-object part correspondence loss (ablation) | LucidPPN · evaluated datasets | top-1 accuracy | Helped |
| Delay the start of ColorNet training (ablation) | LucidPPN · evaluated datasets | top-1 accuracy | No clear effect |
Surprising findings
- On CARS, adding the color branch does not improve over the grayscale ShapeTexNet alone (91.7 vs 91.7), suggesting color is not informative for car classification.
- LucidPPN is slightly less accurate than PIP-Net on CUB and DOGS in the main comparison, despite the interpretability gains.
Limitations
- The architecture only disentangles color; texture, shape, and contrast remain entangled in the ShapeTexNet branch.
- Depends on PDiscoNet part masks, and inherits spatial misalignment and non-obvious PP interpretation limits from PP-based methods.
- Accuracy is comparable to but not uniformly better than baselines (lower on CUB and DOGS).
Experimental setup
Datasets: CUB-200-2011 Stanford Cars Stanford Dogs Oxford 102 Flower
Benchmarks: CUB-200-2011 Stanford Cars Stanford Dogs Oxford 102 Flower
Metrics: top-1 accuracy IoU color sparsity user study accuracy
Baselines: ProtoPNet ProtoTree ProtoPShare ProtoPool PIP-Net
Builds on: ProtoPNet PIP-Net PDiscoNet ConvNeXt
Closest papers by meaning
- Interpretable Image Classification via Non-parametric Part Prototype Learning (CVPR 2025 · similarity 0.91)
- Improving Prototypical Visual Explanations with Reward Reweighing, Reselection, and Retraining (ICML 2024 · similarity 0.87)
- Rashomon Sets for Prototypical-Part Networks: Editing Interpretable Models in Real-Time (CVPR 2025 · similarity 0.87)
- ProtoArgNet: Interpretable Image Classification with Super-Prototypes and Argumentation (AAAI 2025 · similarity 0.86)
- Post-hoc Part-Prototype Networks (ICML 2024 · similarity 0.86)
- Keep the Faith: Faithful Explanations in Convolutional Neural Networks for Case-Based Reasoning (AAAI 2024 · similarity 0.85)
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.