Source-linked AI summary
Spatial-temporal Concept based Explanation of 3D ConvNets
Ying Ji, Yu Wang, Kensaku Mori, Jien Kato
TL;DR
Interpretation of 3D video recognition ConvNets is relatively understudied, while existing attention methods separate spatial and temporal cues and use pixel-level regions. 3D ACE represents videos with supervoxels and scores spatial-temporal concepts; experiments report explanations consistent with human cognition and show how adding or removing concepts affects accuracy.
Problem
Interpretation of 3D video recognition ConvNets is relatively understudied, and existing RNN attention separates spatial and temporal cues while producing pixel-level regions without human-understandable high-level information.
Source: §1 · p. 1Method
3D ACE segments videos into supervoxels, clusters similar segments into spatial-temporal concepts, and scores concept importance using class-logit gradients.
Source: §2 · p. 1, §2.1 · p. 1, §2.2 · p. 2Results
Takeaways & Limitations
Abstract
from arXiv · showhide
Recent studies have achieved outstanding success in explaining 2D image recognition ConvNets. On the other hand, due to the computation cost and complexity of video data, the explanation of 3D video recognition ConvNets is relatively less studied. In this paper, we present a 3D ACE (Automatic Concept-based Explanation) framework for interpreting 3D ConvNets. In our approach: (1) videos are represented using high-level supervoxels, which is straightforward for human to understand; and (2) the interpreting framework estimates a score for each voxel, which reflects its importance in the decision procedure. Experiments show that our method can discover spatial-temporal concepts of different importance-levels, and thus can explore the influence of the concepts on a target task, such as action classification, in-depth. The codes are publicly available.
1. Introduction
Although 2D image ConvNets have concept-based explanations, interpretation of 3D video ConvNets remains limited; existing RNN attention separates spatial and temporal attention and uses pixel-level regions. The paper extends 2D ACE [3] to 3D ACE, with Kinetics experiments reporting explanations consistent with human cognition.
1. Introduction
- For 2D ConvNets, CAVs quantify the importance of user-defined concepts [4], and ACE explores relationships between image segments and class predictions [3].Source: §1 · p. 1
- Existing 3D video interpretation studies are few, mainly because video data are computationally costly and complex.Source: §1 · p. 1
- Existing RNN-based temporal attention separately generates spatial and temporal attention, while its extracted regions remain pixel-level and lack human-understandable high-level information.Source: §1 · p. 1
- 3D ACE extends concept-based explanation to video ConvNets, and qualitative and quantitative Kinetics results are reported as consistent with human cognition.The experiments use popular network architectures.Source: §1 · p. 1
2. Proposed method
3D ACE represents videos through supervoxels, groups similar segments into spatial-temporal concepts, and scores each concept's importance to a class decision. Its procedure combines multiresolution segmentation and feature extraction with concept activation vectors and logit-gradient scores.
2. Proposed method
- 3D ACE segments videos into supervoxels, clusters similar segments within each class into spatial-temporal concepts, and scores each concept for that class.The network pays more attention to concepts with higher scores.Source: §2 · p. 1
2.1. Supervoxel representation
- Each video is segmented at three resolutions to preserve hierarchical information, producing supervoxels of different sizes.Source: §2.1 · p. 1
- To reduce redundant supervoxels, the method retains the most distinguishable ones, then extracts their top-layer features using a 3D ConvNet trained from scratch.Source: §2.1 · p. 1
2.2. Concept-based explanation
- Within each class, similar supervoxels are clustered into concepts, whose CAVs are learned by separating concept segments from random videos with a linear classifier.Source: §2.2 · pp. 1–2
- A concept's importance score uses the gradient of the class logit with respect to video activations along its CAV; the aggregate score is the fraction of class videos positively influenced.Higher aggregate scores mark concepts the network focuses on more strongly.Source: §2.2 · p. 2
3. Experiment
Experiments on 10 randomly selected Kinetics-700 classes test adding and removing concepts and visualize concepts extracted for “bending back.”
3. Experiment
- The experiment uses 10 classes randomly selected from Kinetics-700.Source: §3 · p. 2
- For each model, Table 1 compares adding the highest-score, randomly selected, and lowest-score concepts.The first, second, and third rows correspond to those three groups, respectively.Source: §3 · p. 2, Table 1 · p. 2
- For “bending back,” highest-importance concepts highlight the bending action and body, while lowest-score regions are described as almost meaningless.Source: §3 · p. 2
- After adding 5 concepts, accuracy can reach over 70% of the baseline.The baseline is classification accuracy on the chosen 10-class dataset.Source: Table 1 · p. 2, §3 · p. 2
- Removing the most important concepts obviously decreases classification accuracy.The paper interprets these addition and removal results as revealing which concepts the 3D ConvNets focus on and their role in prediction.Source: Table 2 · p. 2, §3 · p. 2
4. Conclusion
The conclusion presents a spatial-temporal concept-based framework for explaining 3D ConvNets, contrasting its human-understandable high-level explanations with low-level pixel-based methods and reporting effectiveness in extensive experiments.
4. Conclusion
- The paper proposes a spatial-temporal concept-based explanation framework for 3D ConvNets.Source: §4 · p. 2
- The framework provides human-understandable high-level explanations, unlike prior low-level pixel-based methods.Source: §4 · p. 2
- The conclusion reports that extensive experiments show the proposed method’s effectiveness.Source: §4 · p. 2