A Comprehensive Review of YOLO Architectures in Computer Vision: From YOLOv1 to YOLOv8 and YOLO-NAS

A Comprehensive Review of YOLO Architectures in Computer Vision: From YOLOv1 to YOLOv8 and YOLO-NAS Juan Terven Diana‐Margarita Córdova‐Esparza Julio-Alejandro Romero-González

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

A Comprehensive Review of YOLO Architectures in Computer Vision: From YOLOv1 to YOLOv8 and YOLO-NAS 2023 Machine Learning and Knowledge Extraction 10.3390/make5040083 terven2023-a-comprehensive-review-of-yolo-arc terven2023-a-comprehensive-review-of-yolo-arc.txt, the project's staged copy a1a45a86bfc4d708ba2ac7d9b6a458be62900bbf cc-by artificial-intelligence

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them the YOLO family of real-time object detectors from v1 to v8, and the single number the whole family is ranked by Quotation withheld because this record is not cleared for redistribution. what changed architecturally at each YOLO iteration, and on what measure the improvements are quoted Quotation withheld because this record is not cleared for redistribution. a version-by-version account of the architecture changes, prefaced by an explanation of the evaluation metric that shows it is not one metric Quotation withheld because this record is not cleared for redistribution. method Quotation withheld because this record is not cleared for redistribution. real-time object detection, whose published progress is stated almost entirely in points of AP Quotation withheld because this record is not cleared for redistribution. review review Quotation withheld because this record is not cleared for redistribution. AP is built on the precision-recall curve, and precision and recall trade against each other - detecting more objects raises recall and produces more false positives - so the single number is an area under a curve, not a count Quotation withheld because this record is not cleared for redistribution. what counts as a correct detection at all is set by an IoU threshold, and COCO averages over a RANGE of thresholds from 0.5 to 0.95 - so the definition of `correct` is itself a parameter being averaged over Quotation withheld because this record is not cleared for redistribution. TWO NAMES FOR ONE THING: `average precision (AP), traditionally called mean average precision (mAP)`, and `The COCO dataset makes no distinction between AP and mAP`. The review has to declare which word it will use before it can use it Quotation withheld because this record is not cleared for redistribution. AND ONE NAME FOR TWO COMPUTATIONS: `The AP is computed differently in the VOC and in the COCO datasets` - 11-point interpolation over 20 categories at one IoU threshold, against 101-point interpolation over 80 categories averaged over ten IoU thresholds Quotation withheld because this record is not cleared for redistribution. THE CONSEQUENCE IS STATED PLAINLY AND IT IS ABOUT COMPARABILITY: `The differences in AP calculation make it hard to directly compare the performance of object detection models across the two datasets` Quotation withheld because this record is not cleared for redistribution. the improvements the field reports are quoted in that same word - 2.3 points from simOTA, 2.4 from stronger augmentation - so the increments inherit whatever the number means on the day Quotation withheld because this record is not cleared for redistribution. the review derives AP separately for each dataset convention before reporting any model's score, rather than treating the metric as given eliminates: a reader lining up a YOLOv2 number against a YOLOv5 number - the first is VOC AP and the second is COCO AP, and they are different quantities · outcome: held Quotation withheld because this record is not cleared for redistribution. the naming collision is settled explicitly - 'in the rest of this paper, we will refer to this metric as AP' - instead of the two words being used interchangeably eliminates: a review whose own text is ambiguous about which of two names it means, in a document whose whole purpose is to compare numbers · outcome: held Quotation withheld because this record is not cleared for redistribution. whether the reported per-version gains are comparable to one another eliminates: nothing - the gains are quoted in AP points and the review has just shown AP is two computations, so a 2.3-point gain and a 2.4-point gain are only comparable if both were measured under the same convention, which the review does not state per gain · outcome: absent-not-in-staged-text Quotation withheld because this record is not cleared for redistribution. average precision average precision Quotation withheld because this record is not cleared for redistribution. Intersection over Union Intersection over Union Quotation withheld because this record is not cleared for redistribution. precision precision Quotation withheld because this record is not cleared for redistribution. recall recall Quotation withheld because this record is not cleared for redistribution. Non-maximum suppression Non-maximum suppression Quotation withheld because this record is not cleared for redistribution. post-processing technique post-processing technique Quotation withheld because this record is not cleared for redistribution. YOLO YOLO Quotation withheld because this record is not cleared for redistribution. You Only Look Once You Only Look Once Quotation withheld because this record is not cleared for redistribution. single pass single pass Quotation withheld because this record is not cleared for redistribution. two-steps two-steps Quotation withheld because this record is not cleared for redistribution. anchor boxes anchor boxes Quotation withheld because this record is not cleared for redistribution. prior boxes prior boxes Quotation withheld because this record is not cleared for redistribution. Dimension clusters Dimension clusters Quotation withheld because this record is not cleared for redistribution. objectness score objectness score Quotation withheld because this record is not cleared for redistribution. backbone architecture backbone architecture Quotation withheld because this record is not cleared for redistribution. Darknet-19 Darknet-19 Quotation withheld because this record is not cleared for redistribution. YOLO → You Only Look Once YOLO is the acronym that stands for You Only Look Once. Abbreviation expansion is not is-a, part-of, same-as, distinct-from, contrasted-with, operationalised-as, causes, or renamed-across-fields. Quotation withheld because this record is not cleared for redistribution. AP / mAP - THE SAME QUANTITY UNDER TWO NAMES, and the review says so in its own first metrics sentence. `Traditionally called mean average precision`; COCO makes no distinction; the paper picks `AP` by fiat. A FIELD'S CENTRAL METRIC WITH AN UNSETTLED NAME Quotation withheld because this record is not cleared for redistribution. AP, VOC SENSE - 11-point interpolation of the precision-recall curve, averaged over 20 categories, at a single IoU threshold Quotation withheld because this record is not cleared for redistribution. AP, COCO SENSE - 101-point interpolation, 80 categories, averaged over IoU thresholds from 0.5 to 0.95 in steps of 0.05. THE SAME SYMBOL AS THE VOC SENSE AND A DIFFERENT QUANTITY, which is why numbers from the two eras of YOLO cannot be lined up Quotation withheld because this record is not cleared for redistribution. AP50 - the one COCO figure that is a single-threshold measurement, named separately BECAUSE it is the exception. The existence of a special name for the simple case is the clearest sign that the unqualified word is not simple Quotation withheld because this record is not cleared for redistribution. Non-maximum suppression - post-processes multiple overlapping predictions by retaining the most accurate bounding boxes Quotation withheld because this record is not cleared for redistribution. Post-processing technique - NMS is explicitly placed in this method class Quotation withheld because this record is not cleared for redistribution. YOLO - completes object detection through a single network pass rather than proposal generation followed by classification Quotation withheld because this record is not cleared for redistribution. You Only Look Once - the full phrase encoded by the YOLO acronym Quotation withheld because this record is not cleared for redistribution. Single pass - the design choice named by YOLO and contrasted with proposal/classifier pipelines Quotation withheld because this record is not cleared for redistribution. Two-steps - first detect candidate object regions, then classify the region proposals Quotation withheld because this record is not cleared for redistribution. Anchor boxes - predefined shapes that match prototypical object shapes and receive coordinate/class predictions Quotation withheld because this record is not cleared for redistribution. Prior boxes - the review's co-occurring name for anchor boxes Quotation withheld because this record is not cleared for redistribution. Dimension clusters - k-means-derived box priors balancing recall against model complexity Quotation withheld because this record is not cleared for redistribution. Objectness score - YOLOv3 logistic score set to one for the anchor with highest ground-truth overlap Quotation withheld because this record is not cleared for redistribution. Backbone architecture - YOLOv2's feature extractor, called Darknet-19 Quotation withheld because this record is not cleared for redistribution. Darknet-19 - a backbone with nineteen convolutional and five max-pooling layers Quotation withheld because this record is not cleared for redistribution. critiques → cross-dataset comparison of object detection performance, since AP means a different computation on VOC and on COCO Quotation withheld because this record is not cleared for redistribution. supports → the COCO AP as the current standard, on the ground that it evaluates localisation at a range of IoU thresholds rather than one Quotation withheld because this record is not cleared for redistribution. Non-maximum suppression —is-a→ post-processing technique Non-maximum suppression post-processing technique the relation itself is separately evidenced · strength of the relation: 1 · note: The definition explicitly states that NMS is a post-processing technique. Quotation withheld because this record is not cleared for redistribution. single pass —contrasted-with→ two-steps single pass two-steps the relation itself is separately evidenced · strength of the relation: 1 · note: The review contrasts the single network pass with earlier two-step proposal/classifier methods using 'as opposed to'. Quotation withheld because this record is not cleared for redistribution. backbone architecture —same-as→ Darknet-19 backbone architecture Darknet-19 the relation itself is separately evidenced · strength of the relation: 1 · note: The review says the YOLOv2 backbone architecture is called Darknet-19. Quotation withheld because this record is not cleared for redistribution. TWO NAMES FOR ONE THING: AP is `traditionally called mean average precision (mAP)`, and the COCO dataset makes no distinction between them - so the review declares which word it will use before using it TWO NAMES FOR ONE THING: AP is `traditionally called mean average precision (mAP)`, and the COCO dataset makes no distinction between them - so the review declares which word it will use before using it Quotation withheld because this record is not cleared for redistribution. AND ONE NAME FOR TWO COMPUTATIONS: THE AP IS COMPUTED DIFFERENTLY IN THE VOC AND IN THE COCO DATASETS AND ONE NAME FOR TWO COMPUTATIONS: THE AP IS COMPUTED DIFFERENTLY IN THE VOC AND IN THE COCO DATASETS Quotation withheld because this record is not cleared for redistribution. COCO has 80 categories and uses 101-point interpolation, averaging AP over multiple IoU values rather than one, except AP50 COCO has 80 categories and uses 101-point interpolation, averaging AP over multiple IoU values rather than one, except AP50 Quotation withheld because this record is not cleared for redistribution. differences in AP calculation make direct comparison of object detection models across the two datasets hard differences in AP calculation make direct comparison of object detection models across the two datasets hard Quotation withheld because this record is not cleared for redistribution. simOTA, the simplified version proposed by YOLOX, increased AP by 2.3 points simOTA, the simplified version proposed by YOLOX, increased AP by 2.3 points Quotation withheld because this record is not cleared for redistribution. the commonly used metric for evaluating object detection, measuring average precision across all categories - traditionally called mean average precision, and COCO makes no distinction between the two average precision Quotation withheld because this record is not cleared for redistribution. the overlap between the ground truth and predicted bounding boxes, thresholded to decide whether a prediction is a true positive Intersection over Union Quotation withheld because this record is not cleared for redistribution. recall thresholds in the COCO interpolation, against 11 in VOC - and 80 categories against 20 101 value source note: THE DESCRIPTION DRAWS ON THE PAPER BEYOND THIS QUOTE: the VOC's 20 categories are stated just before ('VOC Dataset This dataset includes 20 object categories.'). Quotation withheld because this record is not cleared for redistribution. the single IoU threshold behind AP50, the one COCO figure that is not averaged over a range 0.5 Quotation withheld because this record is not cleared for redistribution. average precision average precision Quotation withheld because this record is not cleared for redistribution. precision precision Quotation withheld because this record is not cleared for redistribution. recall recall Quotation withheld because this record is not cleared for redistribution. confidence threshold confidence threshold Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://doi.org/10.3390/make5040083.