Open world object detection applied in semantic scene construction

Large-scale visual models have dramatically improved object detection speed and accuracy, but they still face challenges like semantic confusion and label errors when deployed in specific domains. To address these issues, we introduce MMRPA (Modified Multi-Scale Random Prim Algorithm), a novel approach that integrates Prim algorithm into the proposal generation mechanism. By leveraging a classifier that distinguishes between foreground and background, our method ensures the model develops an unbiased feature representation, enhancing the detection of unknown classes. We propose a random proposal generation strategy, which offers flexible candidate locations and uses a multi-scale random prim algorithm to establish constraints that help filter candidate boxes. This randomness is independent of known object distributions, preventing overfitting to specific locations of known objects during training. Furthermore, a lightweight knowledge-distillation strategy enables the model to learn emerging categories from very limited annotations while mitigating catastrophic forgetting. Extensive experiments on Pascal-VOC and MS-COCO demonstrate that MMRPA substantially improves both known and unknown object detection. On VOC, MMRPA achieves 68.7%, outperforming the baseline MDETR by 13.7%relative gain in mAP (mean Average Precision). On COCO, MMRPA achieves an UR of 28.3%, surpassing CAT by 4.6% and outperforming OW-DETR by a factor of 2.8. Code is available at https://github.com/YXblank/unknown .