TL;DR We propose a scalable 2D-to-3D lifting pipeline that converts single-view images into scale-calibrated 3D (depth, camera poses, point clouds, pseudo-RGBD), releasing COCO-3D and Objects365-v2-3D datasets. Pretraining on these data boosts 3D perception, enables zero-shot 3D segmentation, and strengthens 3D vision-language/MLLM spatial reasoning — greatly reducing reliance on costly sensor-collected 3D data.

TL;DR Rasterization can offer complementary benefits to map vectorization. We propose (i) MapVR, a precise map vectorization framework bridging rasterization and vectorization, and (ii) a highly sensitive rasterization-based metric for map vectorization.

TL;DR IMFA (Iterative Multi-scale Feature Aggregation) is the first generic paradigm to efficiently leverage multi-scale features in Transformer-based detectors. It only samples multi-scale features from a few crucial keypoints within a few promising regions, achieving high accuracy at small computational overheads.

TL;DR Meta-DETR performs image-level meta-learning-based prediction and exploits inter-class correlation to enhance generalization to new classes, entirely bypassing the proposal quality gap between base and novel classes.

TL;DR SAM-DETR++ extends the semantics-aligned matching mechanism to fuse multi-scale features that are inherently unaligned in semantics, achieving even faster convergence and superior detection accuracy.

TL;DR SAM-DETR converges within 12 epochs and outperforms the strong Faster R-CNN (w/ FPN) baseline on COCO, using the proposed semantic-aligned matching mechanism.

TL;DR Defect-GAN performs high-fidelity defect synthesis with many normal samples and a limited number of defect samples. The technique is patent-protected and is being used in real scenarios.

TL;DR A fine-tuning-based few-shot object detector that learns new concepts without forgetting via weight-normalized sub-networks.

TL;DR Realistic facial expression editing via separate branches focusing on certain areas (eyes, noses, etc.), performed iteratively.

TL;DR One of the earliest works to apply Faster R-CNN to rotated object detection in remote sensing images, using multi-level contextual information.