anchor box
In plain wordsA 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.
One of a fixed set of reference bounding boxesbounding box, n.: The rectangle a vision system draws around an object it has found, stored as a handful of numbers giving the rectangle's position and size. with preset scales and aspect ratios, tiled at every position of a detector's feature mapfeature map, n.: The grid of numbers produced when one learned pattern detector is scanned across an image, showing where, and how strongly, that pattern appears.. For each anchor the network predicts whether it covers an object and how to shift and resize it to fit (box regression); in training, anchors are assigned to ground-truth objects by intersection over unionintersection over union, n.: A score from 0 to 1 for how well two shapes overlap, such as a predicted box and the true box around an object: the shared area divided by the combined area.. Anchors turn detection into dense classification and regression over a regular grid and let one location represent objects of different shapes. Faster R-CNN used nine anchors per location (three scales × three aspect ratios); SSD called them default boxes; YOLOv2 chose anchor shapes by k-means clustering of training boxes. Their drawbacks (many hyperparameters, a large imbalance between the few positive and the many negative anchors, extra computation) motivated anchor-free detectors such as CornerNet, FCOS and CenterNet (2018–2019), which predict objects from points or corners, and query-based detectors such as DETRDETR, prop. n.: An object detector built on the transformer, the design behind modern chatbots, that outputs a short list of objects in one go instead of sorting through thousands of guessed boxes..
Our anchors were too wide for utility poles, so we added a 1:6 aspect ratio.
Origin
Introduced as 'anchors' by Shaoqing Ren, Kaiming He, Ross Girshick and Jian Sun in Faster R-CNN (arXiv June 2015), where they underpin the region proposal network; unlike the box 'priors' that Google's MultiBox detector (2014) obtained by clustering, Faster R-CNN's anchors repeat at every image position. Wei Liu and colleagues' SSD detector (December 2015) used equivalent 'default boxes'.
Usage note
Many real-time detectors went anchor-free from about 2021 (YOLOX; Ultralytics YOLOv8 in 2023), but anchor-based designs remain common in deployed systems.
Related
- 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.
- bounding boxbounding box, n.: The rectangle a vision system draws around an object it has found, stored as a handful of numbers giving the rectangle's position and size.
- R-CNNR-CNN, prop. n.: A family of object detectors that first pick out image regions likely to contain objects, then use a neural network to label each region and tighten its box.
- YOLOYOLO, prop. n.: A family of fast object-detection models that look at an image once and immediately output boxes and labels for everything they recognize, quick enough for live video.
- DETRDETR, prop. n.: An object detector built on the transformer, the design behind modern chatbots, that outputs a short list of objects in one go instead of sorting through thousands of guessed boxes.
- intersection over unionintersection over union, n.: A score from 0 to 1 for how well two shapes overlap, such as a predicted box and the true box around an object: the shared area divided by the combined area.
Sources
- Ren, S., He, K., Girshick, R. & Sun, J. (2015). Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. NeurIPS 2015.
- Liu, W. et al. (2016). SSD: Single Shot MultiBox Detector. ECCV 2016.
Mentioned in
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., bounding boxbounding box, n.: The rectangle a vision system draws around an object it has found, stored as a handful of numbers giving the rectangle's position and size., 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., R-CNNR-CNN, prop. n.: A family of object detectors that first pick out image regions likely to contain objects, then use a neural network to label each region and tighten its box., YOLOYOLO, prop. n.: A family of fast object-detection models that look at an image once and immediately output boxes and labels for everything they recognize, quick enough for live video., AnchorsAnchors, prop. n.: A method that explains a prediction with a simple if-then rule that locks in the result — for example, if an email contains both 'invoice' and 'urgent', the filter almost always flags it., DETRDETR, prop. n.: An object detector built on the transformer, the design behind modern chatbots, that outputs a short list of objects in one go instead of sorting through thousands of guessed boxes.