You Only Look Once: Unified, Real-Time Object Detection
2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016 · paper · AI and machine learning
Presents YOLO, which treats object detection as a single regression from image pixels to bounding boxes and class probabilities. The base model runs at 45 frames per second, a smaller version at 155, with more localization errors but fewer background false positives.
Checked: PDF from arxiv.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
1 finding
Wording Sec. 2.2: "For the first epochs we slowly raise the learning rate from 10−3 to 10−2." The number of warm-up epochs is missing.
- p. 4 Sec. 2.2, learning rate schedule paragraph
For the first epochs we slowly raise the learning rate from 10−3 to 10−2.
- p. 4 Sec. 2.2, preceding paragraph
We train the network for about 135 epochs
Only the authors can say what is right here.
Judge's note: Checked the page 4 image: no numeral appears before "epochs". The listed phases already add up to the stated "about 135" epochs, so the warm-up length cannot be worked out. Only the authors can supply the number.
Second check: another model agreed
Report a problem with this finding
15 minor (typos and the like; most checked by the judge only)
Reference or label The text says "Szegedy et al. ... [8]" (Sec. 3, page 5) and "similar to Szegedy et al. [8]" (Sec. 4.1, page 6), but reference [8] is first-authored by "D. Erhan, C. Szegedy, A. Toshev, and D. Anguelov".
- p. 5 Sec. 3, Deep MultiBox
Szegedy et al. train a
- p. 6 Sec. 4.1
Szegedy et al. [8]
- p. 9 Reference [8]
D. Erhan, C. Szegedy, A. Toshev, and D. Anguelov.
- p. 5 Section 3, Deep MultiBox
Szegedy et al. train a convolutional neural network to predict regions of interest [8]
Judge's note: Checked reference [8] on page 9 and both in-text citations on pages 5 and 6. Szegedy is the second author of [8], so the name in the text does not match the cited entry. The cited work itself may be the intended one.
Second check: another model agreed
Report a problem with this finding
Wording Figure 1 graphic lists step "3. Non-max suppression." but the caption says the system "(3) thresholds the resulting detections by the model’s confidence."
- p. 1 Figure 1 graphic
3. Non-max suppression.
- p. 1 Figure 1 caption
(3) thresholds the resulting detections by the model’s confidence.
- p. 1 Figure 1 caption
(3) thresholds the resulting detections by the model’s confidence
Judge's note: Checked the page 1 image. The numbered step 3 in the figure graphic and step (3) in the caption name different operations. Sec. 2.3 says both NMS and thresholding are used, so this is a labelling inconsistency, not a factual error.
Second check: another model agreed
Report a problem with this finding
Wording Eq. (3): the sums run from i=0 to S² and from j=0 to B, which gives S²+1 cells and B+1 boxes rather than S² and B.
- p. 4 Equation (3), summation limits
loss function:
Judge's note: Checked the equation crop. Every sum is written with lower limit 0 and upper limit S² (or B). With 0-based indexing that counts 50 cells and 3 boxes for S=7, B=2. The intended loss is clear and the fix is obvious (upper limits S²−1 and B−1, or start at 1), so this is an indexing slip.
Second check: another model agreed
Report a problem with this finding
Wording Eq. (3): the box-level terms x_i, y_i, w_i, h_i, C_i carry only a cell index i even though they sit inside sums over j, and the indicator 1^noobj_ij is used but never defined (only 1^obj_i and 1^obj_ij are defined).
- p. 4 Equation (3)
denotes that the jth bounding box predictor in cell i is “re-
Only the authors can say what is right here.
Judge's note: Checked the page 4 image. The terms inside the j-sums have no j subscript. The sentence after the equation defines only 1^obj_i and 1^obj_ij. The meaning can be inferred, but the notation is inconsistent as printed.
Second check: another model agreed
Report a problem with this finding
Wording Fast R-CNN (2007+2012) is listed at 70.0 mAP on VOC 2007 in Table 1, but Sec. 4.3 and Table 2 say "The best Fast R-CNN model achieves a mAP of 71.8%"; the paper does not say which model gives 71.8.
- p. 6 Table 1, Fast R-CNN row
70.0
- p. 6 Sec. 4.3 text and Table 2 first row
The best Fast R-CNN model achieves a mAP of 71.8%
Only the authors can say what is right here.
Judge's note: Checked Table 1 and Table 2 on page 6. The 71.8 may be a different Fast R-CNN variant, but neither the text nor the caption identifies it or relates it to the 70.0 in Table 1. Only the authors can say.
Second check: another model agreed
Report a problem with this finding
Wording Sec. 4.4: "On categories like bottle, sheep, and tv/monitor YOLO scores 8-10% lower than R-CNN or Feature Edit". In Table 3 the gaps do not all fall in 8–10 points for any single comparator.
- p. 7 Sec. 4.4
YOLO scores 8-10% lower than R-CNN or Feature Edit
- p. 7 Table 3, YOLO / Feature Edit / R-CNN rows
22.7
Only the authors can say what is right here.
Judge's note: Read the Table 3 crop. YOLO: bottle 22.7, sheep 52.2, tv 50.8. Feature Edit: 33.1, 61.3, 57.8. R-CNN [13]: 27.9, 54.5, 51.6. R-CNN BB: 32.6, 57.8, 54.1. The Feature Edit gaps are 10.4/9.1/7.0; the plain R-CNN gaps are only 5.2/2.3/0.8. Which R-CNN row is meant, and whether the range is approximate, needs the authors.
Second check: another model agreed
Report a problem with this finding
Wording Sec. 4.5: "its AP degrades less than other methods when applied to artwork". In Figure 5b DPM loses fewer AP points than YOLO on both Picasso and People-Art; YOLO degrades less only in relative terms.
- p. 7 Sec. 4.5
its AP degrades less than other methods when applied to artwork
- p. 8 Figure 5b table
43.2
Only the authors can say what is right here.
Judge's note: Read the Figure 5b table. YOLO goes 59.2 → 53.3 (Picasso) and 45 (People-Art); DPM goes 43.2 → 37.8 and 32. The claim holds only for proportional drops, which the text does not say. The same section also says "DPM maintains its AP well".
Second check: another model agreed
Report a problem with this finding
Typo Sec. 2.2: "followed by a average-pooling layer" should read "an average-pooling layer".
- p. 3 Sec. 2.2
followed by a average-pooling layer
Judge's note: Seen in the page 3 crop exactly as quoted.
Second check: none (judge only)
Report a problem with this finding
Typo Figure 3 caption: "reduce the features space from preceding layers" should read "feature space".
- p. 3 Figure 3 caption
reduce the features space from preceding layers
Judge's note: Seen in the Figure 3 caption crop on page 3.
Second check: none (judge only)
Report a problem with this finding
Typo Sec. 3, MultiGrasp: "or predict it’s class" should read "its class".
- p. 5 Sec. 3, MultiGrasp
or predict it’s class
Judge's note: Seen on the page 5 image.
Second check: none (judge only)
Report a problem with this finding
Typo Sec. 3: the run-in heading "Other Fast Detectors" has no terminating period ("Other Fast Detectors Fast and Faster R-CNN focus on"), unlike the other run-in headings.
- p. 5 Sec. 3 run-in heading
Other Fast Detectors Fast and Faster R-CNN focus on
Judge's note: Checked the page 5 image. The bold heading has no period, while "Deformable parts models.", "R-CNN.", "Deep MultiBox.", "OverFeat." and "MultiGrasp." all end with one.
Second check: none (judge only)
Report a problem with this finding
Typo Sec. 4.2 uses "it’s" for the possessive "its" twice: "it’s detections are publicly available" and "13.6% of it’s top detections are false positives".
- p. 6 Sec. 4.2 first paragraph
it’s detections are publicly avail-
- p. 6 Sec. 4.2 last paragraph
it’s top detections are false positives
- p. 6 Section 4.2, last paragraph
13.6% of it’s top detections are false positives
- p. 6 Sec 4.2
of it’s top detections are false positives
Judge's note: Both instances seen on the page 6 image.
Second check: none (judge only)
Report a problem with this finding
Typo Sec. 4.2: "Fast R-CNN makes much fewer localization errors"; with a countable noun, "far fewer" is the standard usage.
- p. 6 Sec 4.2
makes much fewer local-
Judge's note: Seen on the page 6 image. This is a usage point only.
Second check: none (judge only)
Report a problem with this finding
Typo Table 3 caption reads "Fast R-CNN + YOLO is the forth highest scoring method"; 'forth' should be 'fourth'.
- p. 7 Table 3 caption
Fast R-CNN + YOLO is the forth highest scoring method
- p. 7 Table 3 caption
is the forth highest scoring method
Judge's note: The misspelling is visible in the Table 3 caption on page 7 (crop C016-a). It is a plain typo; the ranking claim was not in question.
Second check: none (judge only)
Report a problem with this finding
Typo Sec. 4.3 reads "since we run each model seperately and then combine the results"; 'seperately' should be 'separately'.
- p. 7 Sec. 4.3
we run each model seperately
- p. 7 Sec 4.3, last paragraph
we run each model seperately and
Judge's note: The misspelling is visible in the last paragraph of Sec. 4.3 on page 7 (crop C017-a). It is a plain typo.
Second check: none (judge only)