CogVis: Must Open-Vocabulary Change Detection Perce... | AI Research

CogVis is a framework designed to improve Open-Vocabulary Change Detection (OVCD) in remote sensing by decoupling temporal perception from semantic reasoning. By reformulating the task into a perception-memory-verification paradigm, the authors aim to eliminate redundant computations and unstable results found in existing methods that process each query as a completely new task.

A New Paradigm for Change Detection

Existing OVCD methods typically follow one of two pipelines: Mask-Compare-Identify (M-C-I) or Identify-Mask-Compare (I-M-C). Both approaches couple temporal evidence, semantic discrimination, and region verification into a fixed cascade. This leads to inefficiencies, as these models often re-analyze the entire scene for every new category query.
CogVis, proposed by Zijie Wang, Chen Zhong, and Wei He, draws inspiration from human visual perception. It separates the process into three distinct stages:

• Scene Change Perceptron (SCP): Extracts a category-agnostic change prior from bi-temporal features once per image pair, which is then shared across all subsequent queries.

• Semantic Memory Calibrator (SMC): Uses a memory bank to retrieve and estimate an image-query-specific decision threshold, compensating for variations in scene and category.

• Adaptive Region Filter (ARF): Verifies candidate regions based on learned semantic, temporal, and structural reliability to reject non-semantic disturbances like shadows or seasonal shifts.

Performance and Efficiency

The authors evaluated CogVis across seven benchmarks, including semantic change detection, binary change localization, and building-damage assessment. According to the paper, CogVis achieved state-of-the-art performance on all evaluated datasets. Notably, it improved mIoU on the xBD building-damage assessment benchmark by 8.17 percentage points compared to previous methods.
Beyond accuracy, the framework improves computational efficiency. By sharing scene-level change perception across queries, CogVis avoids repeating temporal analysis, resulting in a 28.50% increase in inference throughput compared to the next-fastest method.

Franklin Analysis: Methodological Impact

The evidence suggests that the primary advantage of CogVis is its modularity. The ablation studies indicate that the components are complementary: the SCP provides the most significant gains in temporal stabilization, while the SMC is critical for category-dependent calibration.
The authors note that the ARF, while effective, is dependent on the quality of the upstream predictions. When used in isolation, it does not provide the same performance benefits as when it is integrated into the full perception-memory-verification pipeline. This confirms that the framework’s success relies on the sequential, decoupled nature of its design rather than any single component acting as a standalone post-processing tool.

Considerations

While CogVis improves throughput, the authors acknowledge that latency still increases as the number of queried categories grows, because the SMC and ARF stages remain query-specific. Additionally, the ARF is designed to filter unreliable proposals but cannot recover regions that were missed during the initial candidate generation phase. The framework relies on frozen foundation models (SAM3 and CLIP) for its visual and semantic features, meaning its performance is inherently tied to the capabilities of these underlying models.