Feature Pyramid Networks for Object Detection
2017 · paper · AI and machine learning
Introduces the Feature Pyramid Network, which builds multi-scale feature maps from a convolutional network's own hierarchy using a top-down path with lateral connections at marginal extra cost. Used in Faster R-CNN, it reports state-of-the-art single-model COCO results at 6 FPS.
Checked: PDF from arxiv.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
2 findings
Reference or label Sec. 5.1.1: "simulates the effect of reusing the pyramidal feature hierarchy (Fig. 1(b))" and "levels on the bottom-up pyramid (Fig. 1(b))" point to panel (b) ("Single feature map"); the pyramidal feature hierarchy is Fig. 1(c).
- p. 5 Sec. 5.1.1, 'How important is top-down enrichment?'
of reusing the pyramidal feature hierarchy (Fig. 1(b)).
- p. 5 Sec. 5.1.1, 'How important is top-down enrichment?'
of reusing the pyramidal feature hierarchy (Fig. 1(b))
- p. 5 Sec. 5.1.1, next paragraph
levels on the bottom-up pyramid (Fig. 1(b))
- p. 1 Figure 1 panel labels
(c) Pyramidal feature hierarchy
Judge's note: Page 1 image: panel (b) is labelled "Single feature map" and panel (c) "Pyramidal feature hierarchy". The caption says "(c) An alternative is to reuse the pyramidal feature hierarchy". Page 5 image, right column: both sentences cite Fig. 1(b) for the multi-level bottom-up pyramid (Table 1(d)).
Second check: another model agreed
Report a problem with this finding
Wording Sec. 6.1 says "our models run at 6 to 7 FPS", but the Table 6 FPN mask-model times (0.15, 0.19, 0.24, 0.25, 0.25 s) correspond to about 4 to 6.7 FPS.
- p. 8 Sec. 6.1, last paragraph
our mod-els run at 6 to 7 FPS
- p. 8 Table 6, time (s) column
0.25
- p. 8 Sec. 6.1, last paragraph
els run at 6 to 7 FPS
- p. 8 Table 6 time (s) column
0.24
Only the authors can say what is right here.
Judge's note: Checked on the page 8 image. The table caption says all FPN runtimes are on the same M40 GPU. Only the single MLP [5x5] row (0.15 s) falls in the 6-7 FPS range. The other rows, including the best (bold) model, are about 4-5 FPS. Only the authors can say which models or which timing the sentence means.
Second check: another model agreed
Report a problem with this finding
7 minor (typos and the like; most checked by the judge only)
Typo Figure 1 caption: "feature maps are indicate by blue outlines" should read "are indicated by".
- p. 1 Figure 1 caption
feature maps are indicate by blue outlines
- p. 6 5.2 implementation
2 image per GPU
- p. 5 lateral connections paragraph
As a results
- p. 1
feature maps are indicate by blue outlines and thicker
Judge's note: Page 1 image, Figure 1 caption, last two lines: the text reads "feature maps are indicate by blue outlines and thicker outlines". This is a verb-form error. The "2 image per GPU" slip (p.6) is recorded under C015, which covers the related number-agreement slips.
Second check: none (judge only)
Report a problem with this finding
Typo Sec. 1: "The principle advantage of featurizing each level" should read "principal advantage".
- p. 1 Sec. 1, right column
The principle advantage of fea-turizing
- p. 1 Introduction, right column
The principle advantage of fea-
Judge's note: Page 1 image, right column, Introduction: the word printed is "principle". The adjective needed here is "principal".
Second check: none (judge only)
Report a problem with this finding
Typo Sec. 2: "Ghiasi et al. [8] present a Laplacian pyramid presentation for FCNs"; "presentation" should presumably be "representation".
- p. 3 Sec. 2, Methods using multiple layers
cian pyramid presentation for FCNs to progressively refine
- p. 3 Related Work, 'Methods using multiple layers'
cian pyramid presentation for FCNs
- p. 3 Related Work
Lapla- cian pyramid presentation for FCNs
Judge's note: Page 3 crop image: the line reads "Lapla-cian pyramid presentation for FCNs to progressively refine segmentation". "Presentation" is evidently a slip for "representation".
Second check: none (judge only)
Report a problem with this finding
Typo Sec. 5.1: "We report results for 100 and 1000 proposals per images" should read "per image".
- p. 5 Sec. 5.1, first paragraph
100 and 1000 proposals per images
- p. 5 Sec 5.1
1000 proposals per images
Judge's note: Page 5 image, left column, Sec. 5.1: the phrase reads "per images", a number-agreement error.
Second check: none (judge only)
Report a problem with this finding
Typo Sec. 5.1.1: "As a results, FPN has an AR1k score 10 points higher" should read "As a result"; also Sec. 5.2: "Each mini-batch involves 2 image per GPU" should read "2 images".
- p. 5 Sec. 5.1.1, 'How important are lateral connections?'
As a results,
- p. 6 5.2 implementation details
2 image per GPU
Judge's note: Page 5 image, right column, lateral connections paragraph: "As a results,". Page 6 crop, Implementation details: "involves 2 image per GPU". Both are number-agreement slips.
Second check: none (judge only)
Report a problem with this finding
Typo Sec. 5.2 Implementation details reads "Each mini-batch involves 2 image per GPU and 512 RoIs per image"; this should be "2 images per GPU".
- p. 6 Sec. 5.2, Implementation details
volves 2 image per GPU and 512 RoIs per image
- p. 6 Sec 5.2 implementation details
involves 2 image per GPU
- p. 6
involves 2 image per GPU and 512 RoIs per image
Judge's note: The crop of page 6 shows the text exactly as quoted. It is a number-agreement typo.
Second check: none (judge only)
Report a problem with this finding
Typo Reference [27] has "P. Dollar" without the accent, while [14] and [28] have "Dollár"; reference [15] ends "ECCV. 2014." with a period where the other entries use a comma.
- p. 10 Reference [27]
P. Dollar. Learning to seg-
- p. 9 Reference [15]
ECCV. 2014.
Judge's note: Both crops confirm this on the image, so it is not an extraction artefact. [28] on the same page shows "Dollár", and the other entries use the form "In VENUE, year". Cosmetic inconsistency.
Second check: none (judge only)