SWAG-V: Explanations for Video Using Superpixels Weighted by Average Gradients

Thomas Hartley, Kirill Sidorov, Christopher Willis, David MarshallWACV 2022

TL;DR

This paper introduces SWAG-V, a novel method for explaining the predictions of video-based Convolutional Neural Networks (CNNs). By extending the image-based SWAG technique to handle the temporal dimension of videos, SWAG-V generates more accurate and interpretable explanations that better pinpoint important regions and actions within video clips.

What Problem Does This Paper Address?

Convolutional Neural Networks (CNNs) are increasingly used for analyzing video data in critical domains like surveillance and healthcare. However, most existing explanation techniques are designed for image-based CNNs and struggle to effectively account for the temporal dimension inherent in video analysis. This gap leaves us with limited tools to understand why a video-based CNN makes a particular prediction, hindering trust and debugging in these important applications. Existing methods often produce overly coarse or noisy explanations, failing to precisely highlight the spatio-temporal regions that drive the network's decision.

Methods

The core contribution is SWAG-V, an extension of the Superpixels Weighted by Average Gradients (SWAG) method. SWAG-V adapts SWAG for video inputs by:

  • 3D Superpixels: Instead of 2D superpixels for images, SWAG-V generates 3D superpixels that encompass both spatial and temporal dimensions. This allows explanations to track motion and object presence across frames more effectively.
  • Optimized Superpixel Generation: The paper proposes a method to find the optimal number of superpixels. This is crucial because action recognition networks often process smaller spatial dimensions than image networks, and too many superpixels can lead to overly fine-grained, noisy explanations, while too few result in coarse ones. They balance insertion and deletion metrics to find this sweet spot.
  • Combined Image and Gradient Weights (SWAG-VI+G): The authors explore how to best combine image information and gradient information when creating superpixels. Through a grid search, they identify optimal weights (wc and wg) that balance the influence of image features and gradient information for generating informative superpixels.
  • Attribution Method: Guided backpropagation is identified as the most effective attribution method for generating gradients, outperforming baselines and other gradient techniques.

Datasets

  • Kinetics-400: Used for training and evaluating the performance of SWAG-V and other methods on action recognition tasks.
  • UCF101: Used specifically for evaluating the localization abilities of the explanation methods, as it provides bounding box annotations for action classes.

Key Insights and Ideas

  • Temporal Dimension is Key: Simply applying image-based explanation methods to video frames independently is insufficient. Capturing the spatio-temporal nature of actions requires dedicated spatio-temporal explanation techniques.
  • Balancing Act for Explanations: There's a trade-off between "fine" and "coarse" explanations. Too fine, and they become noisy and hard to interpret; too coarse, and they miss crucial details. SWAG-V's optimization for superpixel count and weighting aims to strike this balance.
  • Superpixels as a Bridge: Using superpixels that span both space and time is a powerful way to aggregate gradient information into meaningful, interpretable regions that align with how video models process information.
  • Guided Backpropagation Shines: Among gradient-based attribution methods, guided backpropagation consistently provides the best results for generating informative gradients for video explanation.

Key Results

  • Quantitative Performance: SWAG-V variants (specifically SWAG-VI+G and SWAG-VG) consistently outperform baseline methods and other state-of-the-art techniques (Grad-CAM, Grad-CAM++, Saliency Tubes, EP-3D) on local accuracy metrics (insertion and deletion). SWAG-VI+G is highlighted as the optimal variant, balancing both metrics well.
  • Localization Accuracy: SWAG-V methods significantly outperform others in weak-localization tasks, achieving improvements of around 15-20% over baselines. This indicates SWAG-V's ability to accurately pinpoint actions in both space and time.
  • Efficiency: SWAG-V is significantly more efficient than perturbation-based methods like EP-3D, requiring only a single forward and backward pass compared to EP-3D's thousands of passes. While it has a slight overhead compared to simpler methods, the performance gains justify the cost.
  • Qualitative Results: Visual examples demonstrate SWAG-V's ability to precisely track object motion (e.g., legs in capoeira) and highlight specific actions (e.g., shears in sheep shearing), offering clearer and more interpretable explanations than competing methods.

Limitations and Future Work

  • Computational Overhead: While more efficient than some methods, SWAG-V still introduces computational overhead compared to simpler techniques like guided backpropagation.
  • Parameter Tuning: The optimal parameters (superpixel count, weights) were determined through experimentation. Further research could explore more automated or adaptive methods for parameter selection.
  • Future Work: The authors suggest that their method for optimizing superpixel count could be applied to other explanation techniques that rely on grid-based or superpixel-based explanations. They also imply that further exploration of different spatio-temporal superpixel generation techniques could yield even better results.

Summary updated

BibTeX
@inproceedings{hartley2022swag,
  author = {Thomas Hartley and Kirill Sidorov and Christopher Willis and David Marshall},
  title = {{SWAG-V: Explanations for Video Using Superpixels Weighted by Average Gradients}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2022},
}
Figure