Jack Reilly

Kaicong Huang, Weiheng Oh, Ruimin Ke

Transit video understanding can provide valuable fine-grained data that conventional passenger counters and fare systems cannot capture. However, supervised video models require task-specific annotations, while applying vision-language models (VLMs) directly to long onboard videos is unreliable and costly. To leverage the complementary strengths of both approaches, we propose GHR-VLM, a visual grounded hybrid reasoning framework for zero-shot transit-bus video analytics. It is motivated by the observation that explicit visual grounding can improve VLM reasoning by converting long surveillance streams into compact, passenger-centered spatiotemporal evidence. Specifically, we propose an edge-cloud design in which a lightweight edge-based monitor continuously tracks door status and segments passenger clips. A backend VLM then identifies boarding passengers and classifies payment behavior through a two-stage coarse-to-fine refinement of spatiotemporal evidence. By invoking the VLM only on grounded passenger clips and contact sheets, GHR-VLM reduces cloud inference, avoids payment-specific training data, and supplies the localized evidence that VLMs otherwise struggle to identify. Evaluation on 486 minutes of real-world bus surveillance video demonstrates the potential of grounded edge-cloud reasoning for passenger-level payment analytics while highlighting the challenges posed by degraded video conditions.

Kaicong Huang, Talha Azfar, Jack M. Reilly et al.

Person re-identification faces two core challenges: precisely locating the foreground target while suppressing background noise and extracting fine-grained features from the target region. Numerous visual-only approaches address these issues by partitioning an image and applying attention modules, yet they rely on costly manual annotations and struggle with complex occlusions. Recent multimodal methods, motivated by CLIP, introduce semantic cues to guide visual understanding. However, they focus solely on foreground information, but overlook the potential value of background cues. Inspired by human perception, we argue that background semantics are as important as the foreground semantics in ReID, as humans tend to eliminate background distractions while focusing on target appearance. Therefore, this paper proposes an end-to-end framework that jointly models foreground and background information within a dual-branch cross-modal feature extraction pipeline. To help the network distinguish between the two domains, we propose an intra-semantic alignment and inter-semantic adversarial learning strategy. Specifically, we align visual and textual features that share the same semantics across domains, while simultaneously penalizing similarity between foreground and background features to enhance the network's discriminative power. This strategy drives the model to actively suppress noisy background regions and enhance attention toward identity-relevant foreground cues. Comprehensive experiments on two holistic and two occluded ReID benchmarks demonstrate the effectiveness and generality of the proposed method, with results that match or surpass those of current state-of-the-art approaches.