Focal Loss for Dense Object Detection
2017 IEEE International Conference on Computer Vision (ICCV), 2017 · paper · AI and machine learning
Argues that extreme foreground-background class imbalance explains why one-stage detectors trail two-stage ones, and proposes the Focal Loss, which down-weights easy examples. The resulting RetinaNet detector reportedly matches one-stage speed while exceeding the accuracy of two-stage detectors.
Checked: PDF from arxiv.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
0 findings
None that passed a second check.
15 minor (typos and the like; most checked by the judge only)
Substance Figure 2 inset gives 'RetinaNet-50-500 32.5 73' but Table 1e row '50 500 32.5 ... 72' gives 72 ms for the same model.
- p. 1 Figure 2 inset table, row RetinaNet-50-500
RetinaNet-50-500 32.5 73
- p. 6 Table 1e, row depth 50 scale 500
50 500 32.5 50.9 34.8 13.9 35.8 46.7 72
- p. 6 Table 1e, row depth 50 scale 500
72
Only the authors can say what is right here.
Judge's note: Checked the Figure 2 inset on p1: RetinaNet-50-500 is 32.5 AP and 73 ms. Table 1e on p6 gives depth 50, scale 500: 32.5 AP and 72 ms. The other inset entries match Table 1e exactly (34.4/90 and 37.8/198), and Table 2's caption points to Table 1e and Figure 2 together for speed versus accuracy. The paper gives no reason for the 1 ms difference. Only the authors know which value is right.
Second check: another model agreed
Report a problem with this finding
Reference or label Reference [9] gives the year 2016 for an arXiv identifier from January 2017: "DSSD: Deconvolutional single shot detector. arXiv:1701.06659, 2016."
- p. 10 Reference [9]
arXiv:1701.06659, 2016
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: arXiv identifiers of the form YYMM.NNNNN encode submission year and month, so 1701 means January 2017. The page 10 image confirms that the entry prints 'arXiv:1701.06659, 2016.' Either the year or the identifier needs checking by the authors.
Second check: another model agreed
Report a problem with this finding
Wording The Figure 2 caption says 'including the best reported Faster R-CNN [28] system from [20]', but Table 2 lists Faster R-CNN w TDM [32] at 36.8 AP, above Faster R-CNN w FPN [20] at 36.2, and §5.3 calls the TDM model 'the top-performing Faster R-CNN model'.
- p. 1 Figure 2 caption
including the best reported Faster R-CNN [28] system from [20]
- p. 8 Table 2 and §5.3
top-performing Faster R-CNN model based on Inception-ResNet-v2-TDM [32]
Judge's note: Checked Table 2 and §5.3 on p8 and the Figure 2 caption on p1. Within the paper's own table, the FPN system from [20] is not the best reported Faster R-CNN. The caption may mean the best Faster R-CNN among the systems plotted, or among ResNet backbones, but it does not say so. This is a wording point.
Second check: another model agreed
Report a problem with this finding
Wording §5.3 says "5.9 point AP gap (39.1 vs. 33.2) with the closest competitor, DSSD [9], while also being faster, see Figure 2", but in Figure 2 the RetinaNet-101-800 model takes 198 ms and DSSD513 takes 156 ms.
- p. 8 Section 5.3
5.9 point AP gap (39.1 vs. 33.2) with the closest competitor, DSSD [9], while also being faster, see Figure 2
- p. 1 Figure 2 inset table
[F] DSSD513 [9]
Only the authors can say what is right here.
Judge's note: Figure 2 inset (page 1) lists DSSD513 at 33.2 AP and 156 ms, and RetinaNet-101-800 at 37.8 AP and 198 ms. §5.3 says the 39.1 AP result comes from that same RetinaNet-101-800 architecture, trained longer with scale jitter, so its inference time should be the same 198 ms. 'Faster' does hold for other RetinaNet operating points: RetinaNet-101-500 gets 34.4 AP at 90 ms, which beats DSSD513 on both accuracy and speed. The sentence pairs 'faster' with the 39.1 vs. 33.2 comparison, though, and that pairing does not hold. This is a wording ambiguity, and the authors could say which operating point they meant.
Second check: another model agreed
Report a problem with this finding
Wording In Table 1d, the rows 'OHEM 128 .7' and 'OHEM 1:3 128 .5' both report exactly 31.1 / 47.2 / 33.2 (AP / AP50 / AP75). Can the authors confirm that these two different configurations really gave identical results?
- p. 6 Table 1d, first OHEM row
OHEM 128 .7 31.1 47.2 33.2
- p. 6 Table 1d, first OHEM 1:3 row
OHEM 1:3 128 .5 31.1 47.2 33.2
Only the authors can say what is right here.
Judge's note: Checked Table 1d on p6: the values are identical on all three metrics. The text on p7 discusses only the best OHEM setting (128, .5, 32.8 AP), which this does not affect. The match may be a coincidence or a copying error, and the files cannot settle which.
Second check: another model agreed
Report a problem with this finding
Typo Typos: 'pushes the envelop further' (p1) should be 'envelope', 'we take take our default' (p7) should be 'we take our default', and 'did not shown further gains' (p8) should be 'did not show'.
- p. 1 Introduction, para 3
This paper pushes the envelop further
- p. 1 Introduction
pushes the envelop further
- p. 4 Sec 4
off-the-self
- p. 7 Analysis of the Focal Loss
we take take our default
Judge's note: All three appear on the page images of p1, p7 and p8. The 'off-the-self' item is judged separately under C008.
Second check: none (judge only)
Report a problem with this finding
Typo 'use of features pyramids as in SSD [22] and FPN [20]' should read 'feature pyramids'.
- p. 2 Related Work, last paragraph
use of fea-tures pyramids as in SSD [22] and FPN [20]
Judge's note: The p2 page image, right column, last paragraph, reads 'fea-tures pyramids'.
Second check: none (judge only)
Report a problem with this finding
Typo 'OverFeat [30] was one of the first modern one-stage object detector' should read 'detectors'.
- p. 2 Related Work, One-stage Detectors
was one of the first modern one-stage object detector
Judge's note: Seen on the p2 page image, One-stage Detectors paragraph.
Second check: none (judge only)
Report a problem with this finding
Typo 'is an off-the-self convolutional network' should read 'off-the-shelf'.
- p. 4 §4, first paragraph
an off-the-self
- p. 4 Section 4, RetinaNet Detector
is an off-the-self
- p. 4 Sec 4, first paragraph
off-the-self
Judge's note: Seen on the p4 page image, §4, first paragraph.
Second check: none (judge only)
Report a problem with this finding
Typo 'as well at two-stage detectors such as Fast R-CNN' should read 'as well as'.
- p. 4 §4, FPN Backbone paragraph
as well at two-stage detectors
- p. 4 FPN Backbone paragraph
as well at two-stage
Judge's note: Seen on the p4 page image, FPN Backbone paragraph.
Second check: none (judge only)
Report a problem with this finding
Typo 'This improve AP in our setting' should read 'This improves AP'.
- p. 4 §4, Anchors paragraph
This improve AP in our setting
- p. 4 Anchors paragraph
This improve AP in our setting.
Judge's note: Seen on the p4 page image, Anchors paragraph.
Second check: none (judge only)
Report a problem with this finding
Typo 'The training loss is the sum the focal loss and the standard smooth L1 loss' is missing 'of'.
- p. 6 §4.1 Optimization
The training loss is the sum the focal loss
Judge's note: Seen on the p6 image, §4.1 Optimization.
Second check: none (judge only)
Report a problem with this finding
Typo Duplicated word in §5.1 'Analysis of the Focal Loss': "For this, we take take our default ResNet-101 600-pixel model".
- p. 7 §5.1 Analysis of the Focal Loss
we take take our default
- p. 7
converged model. For this, we take take our default ResNet-
Judge's note: Page 7 image crop shows 'we take take our default ResNet-' plainly; this is a duplicated word.
Second check: code check
Report a problem with this finding
Typo Wrong verb form in §5.2: "increasing beyond 6-9 anchors did not shown further gains" should read "did not show".
- p. 8 §5.2 Anchor Density
not shown further gains
- p. 8 5.2 Anchor Density
did not shown further gains
Judge's note: Page 8 image crop shows 'did not shown further gains' as printed.
Second check: none (judge only)
Report a problem with this finding
Typo 'Derivates' is a misspelling of 'derivatives' in two places: the Figure 6 caption ("Derivates of the loss functions from Figure 5 w.r.t. x") and Appendix B ("For reference, derivates for CE, FL, and FL* w.r.t. x are:"). The Appendix B heading itself reads 'Derivatives'.
- p. 9 Figure 6 caption
Derivates of the loss functions
- p. 9 Appendix B
derivates for CE, FL, and FL∗
- p. 9 Appendix B
derivates for CE, FL, and FL∗w.r.t. x
- p. 9 Figure 6 caption
Derivates of the loss functions from Figure 5 w.r.t. x.
Judge's note: Page 9 images show 'Derivates' in the Figure 6 caption and 'derivates' in the first line of Appendix B, under the heading 'Appendix B: Derivatives'.
Second check: none (judge only)