Deep Learning for Generic Object Detection: A Survey

Deep Learning for Generic Object Detection: A Survey Li Liu Wanli Ouyang Xiaogang Wang Paul Fieguth Jie Chen Xinwang Liu Matti Pietikäinen

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.

Deep Learning for Generic Object Detection: A Survey 2019 International Journal of Computer Vision 10.1007/s11263-019-01247-4 liu2019-deep-learning-for-generic-object-d liu2019-deep-learning-for-generic-object-d.txt, the project's staged copy 34a58b226d9839d380369ebba7e7ed7b9b347068 cc-by computer-science

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 generic object detection with deep learning - what the problem is, how it is scored, and what more than 300 papers have done to it Quotation withheld because this record is not cleared for redistribution. what the field has achieved since deep learning arrived, organised by detection framework, feature representation, proposal generation, context modelling, training strategy and evaluation metric Quotation withheld because this record is not cleared for redistribution. a survey of more than 300 contributions with an explicit account of how detection is evaluated - including that a detection counts as correct only above a chosen overlap threshold, and that the threshold changed Quotation withheld because this record is not cleared for redistribution. critique Quotation withheld because this record is not cleared for redistribution. a field whose measure of success is a threshold and whose object vocabulary is a benchmark's class list - 20, 200 or 91 depending which benchmark Quotation withheld because this record is not cleared for redistribution. review review Quotation withheld because this record is not cleared for redistribution. a detection counts as a true positive when its category matches AND its overlap with the ground-truth box is not smaller than a predefined threshold. Every precision, recall and AP number in the field is downstream of that number - and the number is a convention, not a derivation correct is a threshold Quotation withheld because this record is not cleared for redistribution. PASCAL VOC's mAP is computed at a single IoU of 0.5; MS COCO's is averaged across 0.5 to 0.95 in steps of 0.05. Both are called mAP. A survey of the field has to print a table of which benchmark means which, and any comparison across that boundary is comparing two definitions the same metric name, two quantities Quotation withheld because this record is not cleared for redistribution. 41% of MS COCO objects are small and 24% are large, so size-stratified metrics were introduced. THE AGGREGATE WAS MISLEADING BECAUSE OF WHAT THE DATASET CONTAINED, and the response was to change the measure rather than the data a metric subdivided because the dataset was skewed Quotation withheld because this record is not cleared for redistribution. 20 classes in PASCAL VOC, 200 in ILSVRC, 91 in MS COCO. `Object detection` means detection of whichever of those lists the paper trained against, and the survey says plainly that all of them are much smaller than what a human recognises the object vocabulary is a benchmark's class list Quotation withheld because this record is not cleared for redistribution. one figure panel shows a single class spanning wildly different images; another shows four near-identical images from four different classes. So a detector cannot be evaluated on visual similarity, and the annotation is the only ground truth there is appearance does not determine the label, in either direction Quotation withheld because this record is not cleared for redistribution. THE LONGEST TERM LIST IN THE CORPUS, AND IT IS NOT A LIST OF SYNONYMS - IT IS A LIST OF TERMS WHOSE BOUNDARIES DISAGREE BETWEEN PAPERS, AND IT SITS IN A FOOTNOTE: `Quotation withheld`. Ten terms, attributed, declared unagreed, in a footnote - and the survey then has to pick. Quotation withheld because this record is not cleared for redistribution. whether a detection is correct depends on a predefined overlap threshold rather than on any property of the image Quotation withheld because this record is not cleared for redistribution. mAP means different things under PASCAL VOC and MS COCO, so numbers are not comparable across that boundary Quotation withheld because this record is not cleared for redistribution. the number of object categories in existing benchmark datasets is much smaller than can be recognized by humans - 20, 200 and 91 against a real-world order of 10^4-10^5 Quotation withheld because this record is not cleared for redistribution. the challenge winner is decided by the number of object categories won, with mAP used as a team measure justified by the two rankings having always agreed Quotation withheld because this record is not cleared for redistribution. high accuracy and high efficiency are competing goals, so any single reported figure is a position on a trade-off rather than a level of performance Quotation withheld because this record is not cleared for redistribution. precision and recall from information retrieval, carried into vision as the basis of average precision and then of the field's single headline number Quotation withheld because this record is not cleared for redistribution. MS COCO's average over ten IoU thresholds instead of a single 0.5 eliminates: a detector tuned to clear one overlap value - averaging across 0.5 to 0.95 prices localisation accuracy into the score · outcome: held Quotation withheld because this record is not cleared for redistribution. size-stratified metrics introduced alongside the aggregate, after the object size distribution of the dataset was measured eliminates: a headline number dominated by the size class that happens to be most common - 41% of MS COCO objects are small · outcome: held Quotation withheld because this record is not cleared for redistribution. the challenge's two rankings compared - by mAP and by number of object categories won - before adopting mAP as the team measure eliminates: an aggregate standing in for per-category performance without evidence; the justification is that the two rankings were always the same · outcome: held Quotation withheld because this record is not cleared for redistribution. AP evaluated in a category-specific manner and only then averaged, rather than pooled across categories eliminates: a common category's performance swamping a rare one in the aggregate · outcome: held Quotation withheld because this record is not cleared for redistribution. figure panels showing one class across wildly different appearances and four classes across near-identical appearances outcome (stated by the tagger, no quotation): absent-declared-unreported Quotation withheld because this record is not cleared for redistribution. IOU IOU Quotation withheld because this record is not cleared for redistribution. True Positive True Positive Quotation withheld because this record is not cleared for redistribution. mAP mAP Quotation withheld because this record is not cleared for redistribution. MS COCO MS COCO Quotation withheld because this record is not cleared for redistribution. object categories object categories Quotation withheld because this record is not cleared for redistribution. appearance appearance Quotation withheld because this record is not cleared for redistribution. object class object class Quotation withheld because this record is not cleared for redistribution. intraclass variations intraclass variations Quotation withheld because this record is not cleared for redistribution. accuracy accuracy Quotation withheld because this record is not cleared for redistribution. efficiency efficiency Quotation withheld because this record is not cleared for redistribution. specific instances specific instances Quotation withheld because this record is not cleared for redistribution. broad categories broad categories Quotation withheld because this record is not cleared for redistribution. a correct detection - AS DEFINED: the predicted category equals the ground truth label AND the intersection-over-union with the ground-truth box is not smaller than a predefined threshold Quotation withheld because this record is not cleared for redistribution. mAP - UNDER PASCAL VOC: computed at a single IoU of 0.5 Quotation withheld because this record is not cleared for redistribution. mAP - UNDER MS COCO: averaged across IoU values from 0.5 to 0.95 in steps of 0.05, plus separate metrics for small, medium and large objects Quotation withheld because this record is not cleared for redistribution. object detection - OF SPECIFIC INSTANCES: matching a particular object, such as one face or one tower Quotation withheld because this record is not cleared for redistribution. object detection - OF BROAD CATEGORIES: the generic problem this survey is about, whose category list is whatever the benchmark defines Quotation withheld because this record is not cleared for redistribution. an object class - AS A VISUAL GROUPING: what the four similar-looking images in Fig. 6j would suggest, and which the caption says is wrong Quotation withheld because this record is not cleared for redistribution. an object class - AS A LABEL: a benchmark's annotation, which can hold an astonishing variation in appearance and can separate images that look alike Quotation withheld because this record is not cleared for redistribution. TEN NAMES FOR OVERLAPPING TASKS, AND THE FIELD IS DECLARED TO HAVE NO AGREEMENT ABOUT ANY OF THEM: `there is no universal agreement in the literature on the definitions of various vision subtasks. Terms such as detection, localization, recognition, classification, categorization, verification, identification, annotation, labeling, and understanding are often differently defined`. A survey cannot proceed without saying which it means, and the field supplies ten words and no map. Quotation withheld because this record is not cleared for redistribution. critiques → the benchmark class lists themselves - 20, 200 and 91 categories against a real world of 10^4 to 10^5 Quotation withheld because this record is not cleared for redistribution. qualifies → mAP as a comparable number, by showing that PASCAL VOC and MS COCO compute it at different IoU regimes Quotation withheld because this record is not cleared for redistribution. builds-on → more than 300 research contributions, organised by component of the detection pipeline Quotation withheld because this record is not cleared for redistribution. distinguishes-from → detection of specific instances, which the survey separates out as essentially a matching problem Quotation withheld because this record is not cleared for redistribution. appearance —distinct-from→ object class appearance object class differs in: whether two images look alike or carry the same label · the relation itself is separately evidenced · strength of the relation: 0.95 · differs in (stated by the tagger, no quotation): whether two images look alike or carry the same label · the relation stated: one class holds an astonishing variation in appearance, and four near-identical images belong to four different classes - the label is not recoverable from how things look, in either direction Audit 2026-09-28: DOUBTFUL - read against its span and kept, but the span contrasts intra-class variation with inter-class similarity; distinct-from is loose. Quotation withheld because this record is not cleared for redistribution. intraclass variations —distinct-from→ object categories intraclass variations object categories differs in: whether the difficulty is within one label or between labels · the relation itself is separately evidenced · strength of the relation: 0.95 · differs in (stated by the tagger, no quotation): whether the difficulty is within one label or between labels · the relation stated: the two named sources of detection difficulty - variation among instances sharing a label, and the sheer number of labels to be told apart Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span adds the number of object categories to intraclass variations as a second challenge; distinct-from is loose. Quotation withheld because this record is not cleared for redistribution. specific instances —distinct-from→ broad categories specific instances broad categories differs in: whether the target is one object or a class of them · the relation itself is separately evidenced · strength of the relation: 0.95 · differs in (stated by the tagger, no quotation): whether the target is one object or a class of them · the relation stated: detecting one particular object is essentially a matching problem; detecting a category is the generic task this survey is about Quotation withheld because this record is not cleared for redistribution. accuracy —distinct-from→ efficiency accuracy efficiency differs in: whether the constraint comes from the vision problem or from the deployment device · the relation itself is separately evidenced · strength of the relation: 0.95 · differs in (stated by the tagger, no quotation): whether the constraint comes from the vision problem or from the deployment device · the relation stated: the two competing goals - localising and recognising well enough to tell categories apart, against running in real time within a device's memory Quotation withheld because this record is not cleared for redistribution. a detection counts as correct only if its overlap with the ground truth is not smaller than a predefined threshold a detection counts as correct only if its overlap with the ground truth is not smaller than a predefined threshold Quotation withheld because this record is not cleared for redistribution. MS COCO averages AP across IoU thresholds from 0.5 to 0.95 in steps of 0.05, instead of the traditional single IoU of 0.5. MS COCO averages AP across IoU thresholds from 0.5 to 0.95 in steps of 0.05, instead of the traditional single IoU of 0.5. Quotation withheld because this record is not cleared for redistribution. the benchmark class lists are 20, 200 and 91 categories, much smaller than what humans recognise the benchmark class lists are 20, 200 and 91 categories, much smaller than what humans recognise Quotation withheld because this record is not cleared for redistribution. one object class can span an astonishing variation in appearance while four near-identical images belong to four different classes one object class can span an astonishing variation in appearance while four near-identical images belong to four different classes Quotation withheld because this record is not cleared for redistribution. size-stratified metrics were introduced because 41% of MS COCO objects are small and 24% are large size-stratified metrics were introduced because 41% of MS COCO objects are small and 24% are large Quotation withheld because this record is not cleared for redistribution. the challenge winner is the team that wins the most object categories, with mAP justified as a team measure by the two rankings always agreeing the challenge winner is the team that wins the most object categories, with mAP justified as a team measure by the two rankings always agreeing Quotation withheld because this record is not cleared for redistribution. high accuracy and high efficiency are named as competing goals of the same ideal algorithm high accuracy and high efficiency are named as competing goals of the same ideal algorithm Quotation withheld because this record is not cleared for redistribution. a predicted detection whose category equals the ground-truth label and whose intersection-over-union with the ground-truth box is not smaller than a predefined threshold True Positive Quotation withheld because this record is not cleared for redistribution. average precision computed per category and then averaged over all object categories, adopted as the final measure of performance mAP Quotation withheld because this record is not cleared for redistribution. of the objects in MS COCO are small and 24% are large - the reason size-stratified metrics exist 41% Quotation withheld because this record is not cleared for redistribution. more than this many research contributions are included in the survey 300 Quotation withheld because this record is not cleared for redistribution. mAP mAP Quotation withheld because this record is not cleared for redistribution. IOU IOU Quotation withheld because this record is not cleared for redistribution. object categories object categories Quotation withheld because this record is not cleared for redistribution. IOU IOU 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/10.1007/s11263-019-01247-4.