DETR
An end-to-end object detection model based on the Transformer architecture. DETR replaces conventional anchor boxes and non-maximum suppression with learnable query vectors and a cross-attention mechanism, establishing a foundational paradigm for Transformer-based object detectors.
Timeline (last 90 days)
Empty-annotation images that contain no objects should also be kept in the dataset, which helps the model learn to suppress background false detections.
For a small dataset of 500 images, the recommendation is to first manually annotate 80-100 images precisely as a seed set, fine-tune a lightweight detector with the seed set, then pre-annotate the remaining 400 images, take the union of the two rounds of pre-annotation results, and have humans confirm and correct them.
Mainstream 2D object detection methods can be divided into two-stage detectors (such as Faster R-CNN) and one-stage detectors (such as YOLO, SSD, RetinaNet), as well as the Transformer-based DETR series.
Traditional 2D detectors (such as YOLO and DETR series) are trained on planar images and output bounding boxes in image pixel coordinates, which carry no spherical geometric information.
5.18.0 introduces six breaking changes, including routing FA3 to aiter-flash-attn for gpt-oss on the ROCm platform, DETR image processing acceleration, remapping layer_type for indexers, fixing video input token counting for the vLLM backend, modernizing DINO-series models, and upgrading the Kernels version.
The object query in DETR is a set of randomly initialized learnable embedding vectors, and the lack of spatial prior information is one of the core reasons for its slow convergence.
DETR abandons the anchor box design and non-maximum suppression (NMS) post-processing pipeline found in traditional detectors.
DETR relies on the Hungarian matching algorithm to find the optimal one-to-one correspondence between predicted boxes and ground-truth targets via bipartite graph matching in order to compute the loss.
Facebook AI introduced DETR (DEtection TRansformer) in 2020.
DETR系列模型通过Transformer的集合预测机制消除了传统检测模型（如YOLO）依赖的非极大值抑制（NMS）等后处理步骤，实现端到端检测
All Facts (11)
For a small dataset of 500 images, the recommendation is to first manually annotate 80-100 images precisely as a seed set, fine-tune a lightweight detector with the seed set, then pre-annotate the remaining 400 images, take the union of the two rounds of pre-annotation results, and have humans confirm and correct them.
50%UnverifiedEmpty-annotation images that contain no objects should also be kept in the dataset, which helps the model learn to suppress background false detections.
50%UnverifiedMainstream 2D object detection methods can be divided into two-stage detectors (such as Faster R-CNN) and one-stage detectors (such as YOLO, SSD, RetinaNet), as well as the Transformer-based DETR series.
50%UnverifiedTraditional 2D detectors (such as YOLO and DETR series) are trained on planar images and output bounding boxes in image pixel coordinates, which carry no spherical geometric information.
50%Unverified5.18.0 introduces six breaking changes, including routing FA3 to aiter-flash-attn for gpt-oss on the ROCm platform, DETR image processing acceleration, remapping layer_type for indexers, fixing video input token counting for the vLLM backend, modernizing DINO-series models, and upgrading the Kernels version.
50%UnverifiedThe object query in DETR is a set of randomly initialized learnable embedding vectors, and the lack of spatial prior information is one of the core reasons for its slow convergence.
50%UnverifiedDETR abandons the anchor box design and non-maximum suppression (NMS) post-processing pipeline found in traditional detectors.
50%UnverifiedDETR relies on the Hungarian matching algorithm to find the optimal one-to-one correspondence between predicted boxes and ground-truth targets via bipartite graph matching in order to compute the loss.
50%UnverifiedFacebook AI introduced DETR (DEtection TRansformer) in 2020.
50%UnverifiedDETR系列模型通过Transformer的集合预测机制消除了传统检测模型（如YOLO）依赖的非极大值抑制（NMS）等后处理步骤，实现端到端检测
50%Unverified目标检测传统上由 YOLO、Faster R-CNN、DETR 等专用架构主导，通过专用检测头回归边界框坐标与类别概率
50%