Improving Driving Video Search via Motion

Summary & Methodology Analysis

The TraVEL (Trajectory-Guided Video Embedding Learning) framework addresses the tendency of general-purpose multimodal models to ignore vehicle motion in favor of static scene features. By leveraging Qwen3-VL-Embedding as a base, the authors implement a two-stage training strategy. First, they perform supervised fine-tuning on paired clips and reasoning traces using an InfoNCE objective, which is a contrastive loss function used to pull related data samples together while pushing unrelated ones apart in vector space. This reduces the domain gap between generic vision models and the specific requirements of driving scenarios. The architecture integrates a trajectory-derived reward within Group Relative Policy Optimization (GRPO), a reinforcement learning algorithm that refines model behavior by optimizing policies across groups of outputs to ensure the embedding space accurately reflects vehicle motion while preserving text-to-video alignment. This approach sidesteps the need for heavy perception stacks during inference, focusing instead on optimizing the embedding retrieval performance. Despite these gains, the current evaluation is constrained to a subset of the nuReasoning dataset with a limited retrieval pool. Furthermore, the researchers note that distinguishing between fine-grained maneuvers, such as lane changes versus subtle within-lane shifts, remains a challenge when relying strictly on ego-motion signals.