電腦視覺
CogVis Shares Change Priors Across Remote-Sensing Images, Cutting Latency for Ten-Category Queries to 10.03 Seconds
CogVis decouples cross-temporal image comparison from text-category classification, allowing multiple open-vocabulary queries to reuse the same scene-change features. The authors report state-of-the-art results across seven benchmarks and a 28.5% throughput gain, but did not release the model weights with the code.
Open-vocabulary change detection must answer not only “where did change occur?” but also whether the change matches a text category entered by the user on the fly. Existing pipelines often compare two satellite or aerial images again for every query, coupling temporal differences, text semantics, and region filtering. As the number of query categories grows, the same visual computations are repeated.
CogVis divides the pipeline into three layers. The Scene Change Perceptron first extracts a category-agnostic change prior from frozen bi-temporal features, computing it only once for each image pair. The Semantic Memory Calibrator then estimates a dynamic decision threshold based on the images and text query, correcting score-scale differences across categories. Finally, the Adaptive Region Filter combines semantic, temporal, and regional-structure reliability to remove candidate regions caused by shadows, seasonal differences, and residual registration errors. The underlying SAM3 and CLIP ViT-B/16 models remain frozen, with only three lightweight adapters trained.
The authors report state-of-the-art results on SECOND, SCSCD, CLCD, DSIFN, LEVIR-CD, WHU-CD, and the xBD disaster-damage dataset. The gains are uneven: on LEVIR-CD and WHU-CD, CogVis exceeds the previous best results by only 0.05 and 0.14 points, respectively, while mIoU on xBD rises from 24.51 to 32.68. In tests using an RTX 3090 with 512×512 inputs, CogVis reaches 0.230 FPS with peak memory usage of 4.82 GiB, delivering 28.5% higher throughput than the next-fastest method. Querying ten categories takes 10.03 seconds, compared with 12.56 to 956.60 seconds for per-category baselines. Latency still increases with the number of categories because the calibrator and region filter are not yet shared across queries.
The repository includes training, inference, dry-run, and seven benchmark configurations, enabling engineering teams to inspect the data interfaces and reproduce the workflow. However, the three adapters and the SAM3 and CLIP checkpoints are not included in Git, and the datasets must be prepared separately. The performance comparison is also based on a single hardware configuration and baselines selected by the authors. Future evaluation should examine whether the complete weights are released, whether calibration remains effective across sensors and resolutions, and whether the shared prior retains a linear cost advantage at high volumes of text queries.