Abstract
Video Question Answering (VideoQA) models often achieve high performance by exploiting spurious correlations rather than genuine visual understanding. This paper introduces EIGV (Equivariant and Invariant Grounding), a framework designed to transform "black-box" VideoQA models into self-interpretable systems. By partitioning video content into causal scenes and environmental contexts, EIGV applies causal principles—specifically equivariance and invariance—to ensure models focus on relevant visual cues. Experimental results across major benchmarks like MSVD-QA, MSRVTT-QA, and NExT-QA demonstrate that EIGV not only improves state-of-the-art accuracy but also provides faithful, intrinsic rationales for its predictions.
Introduction
Most current VideoQA models are opaque, making it difficult to discern if a correct answer stems from actual visual reasoning or dataset bias. For instance, a model might answer "riding a horse" correctly not because it identified the action, but because it associated a green meadow background with that specific activity.
Existing post-hoc explainability methods, such as heatmaps, often lack "faithfulness," showing what humans expect rather than what the model actually utilized. To address this, the authors propose intrinsic interpretability: building the explanation mechanism directly into the model's reasoning process so that the rationale—the specific video clips used—is a fundamental component of the final answer.
Methodology
The authors frame VideoQA through the lens of Causal Theory, partitioning a video into two distinct parts:
- Causal Scene ($C$): Frames or objects critical to answering the question.
- Environment Scene ($E$): Irrelevant background or context.
The EIGV Framework
The EIGV framework is a model-agnostic "wrapper" consisting of three primary modules added to a standard VideoQA backbone:
- Grounding Indicator: Employs a Gumbel-Softmax trick to make a "hard" choice, splitting the video into causal and environment segments based on the question.
- Intervener: The core innovation of the framework. It performs two types of interventions:
- Causal-Equivariance: It mixes the causal scene and question with another sample. If the input changes semantically (e.g., mixing "girl riding" with "man surfing"), the predicted answer must change proportionally (e.g., "riding + surfing").
- Environment-Invariance: It swaps the environment scene with random noise or background from other videos. The model is trained to ensure the answer remains unchanged, proving it does not rely on spurious correlations.
- Disruptor: Utilizes contrastive learning to further sharpen the model's ability to distinguish between a helpful rationale and a misleading one.
Experimental Evaluation
The researchers evaluated EIGV on three major benchmarks: MSVD-QA (~50k pairs), MSRVTT-QA (~243k pairs), and NExT-QA (~47.7k pairs). The latter is specifically designed to test causal and temporal reasoning, such as "Why did X happen?", which presents a significant challenge compared to simple descriptive tasks.
Key Results
- State-of-the-Art Performance: EIGV consistently outperformed existing models across all three datasets, with accuracy gains ranging from +1.2% to +2.3% over previous benchmarks.
- Backbone Agnosticism: When integrated with existing architectures like Co-Mem (memory-based), HGA (graph-based), or MSPAN (hierarchy-based), EIGV provided significant boosts, including a +5.2% gain in accuracy for specific backbones.
- Qualitative Grounding: Visualizations confirm that EIGV successfully "zeros in" on correct temporal windows. For example, when queried about a specific action, the model correctly ignores irrelevant segments of the video to focus solely on the causal frames.
Discussion and Future Work
EIGV demonstrates that interpretability and accuracy are not a trade-off; by forcing the model to ignore environmental "noise," it becomes more robust. However, some limitations remain. The current grounding operates at a temporal level (clips); future work could extend this to finer-grained spatial grounding (specific pixels or objects). Additionally, in highly complex scenes with overlapping actions, the model occasionally captures only partial rationales. While the intervention and disruption phases add complexity during training, the framework remains efficient during inference as these modules are removed.
Summary updated