Interpretable Point-Cloud Classification using Multiple Instance Learning

ICCV 2025 Highlight

Three‑dimensional point clouds are ubiquitous in science and industry, yet many state‑of‑the‑art classifiers behave as opaque black boxes. PointMIL introduces an inherently interpretable framework that leverages multiple instance learning (MIL) to jointly improve accuracy and provide fine‑grained, point‑level explanations.

🔮 Overview: We unify common point cloud backbones and feature aggregation strategies within a MIL framework to enable inherently interpretable classification.
💡 Novel dataset: We introduce ATLAS-1, a novel dataset of drug-treated cancer cells, and use it to show how PointMIL captures fine-grained morphological effects of treatment, with implications for drug discovery and mechanism-of-action prediction.
💎 Results & Capabilities: Achieve state-of-the-art performance while offering point-level explanations without the need for post-hoc saliency or perturbation-based methods.
Matt De Vries1,2,3 Reed Naidoo1,2 Olga Fourkioti2 Lucas G. Dent4 Nathan Curry3 Christopher Dunsby3 Chris Bakal1,2
1Sentinal4D   2Institute of Cancer Research   3Imperial College London   4University College London
Paper Code Interactive demo ATLAS-1 Video Summary BibTeX
PointMIL overview

Interactive demo

Load a sample point cloud and visualise PointMIL's point-level importance (gray → red). Toggle class, palette, and thresholds. Drop your own .json files that follow the schema below.

Controls

JSON schema { points: [[x,y,z],...], classes: ["..."], scores: { "Class A": [0..1 per point], ... } }

PointMIL in a nutshell

🔮
Method
Overview
💡
Key
Innovation
🛠️
Algorithm
Details
💎
Results &
Capabilities

🔮 Method Overview

PointMIL is an inherently interpretable framework for point-cloud classification. We treat each point cloud as a bag of instances and learn to aggregate evidence with Multiple Instance Learning (MIL). Unlike post-hoc saliency, PointMIL’s pooling mechanism yields point-level importance as part of the forward pass, so explanations are aligned with the model’s actual decision rule.

PointMIL overview
Figure 1. PointMIL wraps common point backbones with MIL pooling to produce class predictions and point-level importance simultaneously.
  • Backbones: works with standard point-based classifiers (e.g., PointNet, DGCNN, CurveNet).
  • MIL Pooling: Instance / Attention / Additive / Conjunctive variants.
  • Outputs: class probabilities + calibrated point-level importance per class.

💡 Key Innovation

We unify popular point-cloud feature extractors with MIL pooling to make interpretability structural, not post-hoc. The pooling layer acts as a transparent evidence combiner, so the same scores that drive the prediction also surface the discriminative regions.

  • Inherent interpretation: importance = contribution used in the decision, no surrogate maps.
  • Pooling family: swap-in choices (Instance, Attention, Additive, Conjunctive) balance sparsity, stability, and context-sharing.

🛠️ Algorithm Details

Let \( \mathbf{P}=\{\mathbf{p}_i\}_{i=1}^N \subset \mathbb{R}^3 \) with features \( \mathbf{F}=\{\mathbf{f}_i\}_{i=1}^N \in \mathbb{R}^{d_{\text{in}}} \). A point encoder \( f_{\text{enc}} \) yields per-point embeddings \( \mathbf{z}_i \in \mathbb{R}^d \). An MIL head \( f_{\text{MIL}} \) aggregates them into bag-level logits \( \hat{\mathbf{Y}} = f_{\text{MIL}}(\{\mathbf{z}_i\}_{i=1}^N) \in \mathbb{R}^c \), where \( f_{\text{MIL}} \in \{\text{Instance},\text{Attention},\text{Additive},\text{Conjunctive}\} \).

Instance pooling

Classify each point, then average the per-point predictions to get the bag prediction. This assigns per-point responsibility transparently.

\[ \hat{\mathbf{y}}_{i} \in \mathbb{R}^{c} = f_{\mathrm{clf}}(\mathbf{z}_{i}), \qquad \hat{\mathbf{Y}} = \frac{1}{N}\sum_{i=1}^{N} \hat{\mathbf{y}}_{i}. \]

Attention pooling

Learn attention weights over points, pool a weighted feature, then classify the pooled feature.

\[ a_i \in [0,1] = f_{\mathrm{attn}}(\mathbf{z}_i), \qquad \hat{\mathbf{Y}} = f_{\mathrm{clf}}\!\left(\frac{1}{N}\sum_{i=1}^{N} a_i \,\mathbf{z}_i\right). \]

Additive pooling

Weight each point’s features, classify each weighted point, then average the per-point predictions.

\[ a_i \in [0,1] = f_{\mathrm{attn}}(\mathbf{z}_i), \quad \hat{\mathbf{y}}_{i} = f_{\mathrm{clf}}(a_i\,\mathbf{z}_i), \quad \hat{\mathbf{Y}} = \frac{1}{N}\sum_{i=1}^{N} \hat{\mathbf{y}}_{i}. \]

Conjunctive pooling

Learn attention and classification heads independently on features; combine by a weighted mean of per-point predictions.

\[ a_i \in [0,1] = f_{\mathrm{attn}}(\mathbf{z}_i), \quad \hat{\mathbf{y}}_{i} = f_{\mathrm{clf}}(\mathbf{z}_i), \quad \hat{\mathbf{Y}} = \frac{1}{N}\sum_{i=1}^{N} a_i\,\hat{\mathbf{y}}_{i}. \]

Contextual attention (smoothing)

Attention can be sparse; we smooth weights by averaging over each point’s local neighborhood, improving coverage of discriminative regions.

\[ a_i \in [0,1] = f_{\mathrm{attn}}(\mathbf{z}_i), \qquad a_i^{\text{new}} = \frac{1}{k}\!\!\sum_{j \in \mathcal{N}(\mathbf{p}_i)}\! a_j . \]

Interpretability

For Instance, point-level scores come directly from per-point logits: \( \{\hat{\mathbf{y}}_i\}_{i=1}^N \). Additive and Conjunctive scale these by attention: \( \{ a_i\,\hat{\mathbf{y}}_i \}_{i=1}^N \). We apply softmax over classes and read the target class component to get a scalar per point. For Attention, the weights \( \mathbf{a}=\{a_i\}_{i=1}^N \) provide non–class-specific importance.

💎 Results & Capabilities

Interpretability

PointMIL delivers competitive or state-of-the-art accuracy while providing faithful point-level explanations on datasets such as IntrA, RBC, ModelNet40, and drug-response 3D phenotypes.

Interpretation results
Figure 2. PointMIL, CLAIM and PSM interpretability visualisations and corresponding perturbation curves using the Transformer backbone for example-cells from the IntrA dataset. Red points mark regions most responsible for the predicted label.
Clssification results
Figure 3. Classification results on IntrA, RBC, and ModelNet40. All results are shown without a voting strategy on 1024 points. The highest results are shown in bold. Differences between backbones and POINTMIL are shown in violet. Adapted architectures without farthest point sampling results are shown with a †.
Robustness to noise
Figure 4. Interpretability visualisations of PointMIL on a Airplane from ModelNet40 after adding a number (shown on the heading) of noisy points. PointMIL can still focus on salient shape motifs, ignoring noise.
  • Faithful: explanations are produced by the decision mechanism itself.
  • Granular: per-class importance enables side-by-side comparisons across hypotheses.
  • Practical: simple thresholds produce clean, reviewer-friendly overlays for QA and error analysis.

Usage in the Demo

The web demo loads a point cloud, selects a class, and shades points by MIL importance (gray→red). Use the “High: Spheres + shadows” mode to match our publication figures and rotate the scene interactively.

✅ Conclusion

PointMIL makes point-cloud classification interpretable by design, unifying common backbones with transparent MIL pooling. The same mechanism that predicts the class reveals where the model looked, enabling reliable inspection, error analysis, and communication in scientific and industrial settings.

Citation

@inproceedings{DeVries2025Interpretable, title={Interpretable Point Cloud Classification using Multiple Instance Learning}, author={{De Vries}, Matt and Naidoo, Reed and Fourkioti, Olga and Dent, Lucas and Curry, Nathan and Dunsby, Cristopher and Bakal, Chris}, booktitle={International Conference on Computer Vision (ICCV)}, year={2025} }