DETR

'DEtection TRansformer': an object detectionobject detection, n.: Finding every object of interest in a picture or video frame, saying what each one is and drawing a box around it, for example all the cars, cyclists and pedestrians in a street scene. architecture that predicts an image's objects all at once as an unordered set rather than scoring thousands of candidate boxes. A CNNconvolutional neural network, n.: A kind of neural network built for images and other grid-like data. It slides small pattern detectors across the input, finding edges first, then shapes, then whole objects. backbonebackbone, n.: The main, usually pretrained part of a vision model that turns an image into useful internal features; smaller task-specific parts are attached to it, like limbs to a spine. produces image features, a transformertransformer, n.: The design behind today's chatbots and much of modern AI. It handles text by letting every word look at every other word and decide which of them matter for understanding it. encoder relates them across the whole image, and a decoder turns a fixed number (originally 100) of learned 'object queries' into box-and-class predictions in parallel, each query either claiming one object or predicting 'no object'. Training uses a set loss: the Hungarian algorithm finds the best one-to-one (bipartite) matching between predictions and ground-truth objects, and only matched pairs incur class and box losses. Since each object is claimed by exactly one prediction, neither anchorsanchor box, n.: A ready-made box shape, placed at many spots across an image, that an object detector stretches and shifts to fit real objects instead of drawing boxes from scratch. nor non-maximum suppressionnon-maximum suppression, n.: A clean-up step in object detection: when the system draws several overlapping boxes around the same object, it keeps the most confident one and throws away the near-duplicates. are needed, and the design carries over to panoptic segmentationpanoptic segmentation, n.: A complete labeling of an image in which every pixel gets a category, such as road or sky, and every countable object, such as each car or person, also gets its own identity.. The original model matched Faster R-CNN on COCOCOCO, prop. n.: A widely used collection of everyday photos in which people outlined and named the objects, described each picture in sentences and marked body joints; it is used to train and compare vision systems. but needed about 500 training epochs and did poorly on small objects; Deformable DETR (2020), DINO (2022) and the real-time RT-DETR (2023) addressed these weaknesses, and query-based set prediction now underlies many detection, segmentation and open-vocabularyopen-vocabulary detection, n.: Object detection where you type what to look for, such as 'red umbrella' or 'forklift', and the system finds it, even if it was never trained on that exact category. models.

Switching to a DETR-style head let us delete the NMS post-processing from the export script.