Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation
2014 IEEE Conference on Computer Vision and Pattern Recognition, 2014 · paper · AI and machine learning
Introduces R-CNN, which applies large convolutional networks to bottom-up region proposals, with supervised pre-training and fine-tuning. It reports 53.3% mean average precision on PASCAL VOC 2012, over 30% relative improvement, and outperforms OverFeat on ILSVRC2013 detection.
Checked: PDF from arxiv.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
1 finding
Substance full+fg R-CNN fc7 mean segmentation accuracy on VOC 2011 val is '45.8' in Table 5 but '45.7' in Table 7. Table 7's 21 per-category values average to 45.65.
- p. 10 Table 5, full+fg R-CNN fc7 column
45.8
- p. 14 Table 7, row 'full+fg R-CNN fc7', mean column
45.7
- p. 10 Table 5, full+fg R-CNN fc7 column
47.9 45.8
- p. 14 Table 7, full+fg R-CNN fc7 row, mean
full+fg R-CNN fc7 82.3 56.7 20.6 ... 55.2 45.7
Judge's note: Both tables are VOC 2011 val. The text says Appendix E (Table 7) gives the complete per-category results behind Table 5. All other entries match (46.4, 43.0, 42.5, 43.7, 42.1, 47.9). I checked the convention on the fc6 row: its 21 values (bg plus 20 classes) average to 47.87, which matches the printed 47.9. The fc7 row sums to 958.7, giving 45.65. Even allowing ±0.05 rounding on each of the 21 entries, the mean cannot reach 45.75, so Table 5's 45.8 is inconsistent with Table 7.
Second check: another model agreed
Report a problem with this finding
12 minor (typos and the like; most checked by the judge only)
Reference or label Sec. 3.1 credits 'Zeiler and Fergus ... deconvolutional approach in [42]', but [42] is 'M. Zeiler, G. Taylor, and R. Fergus. Adaptive deconvolutional networks for mid and high level feature learning. In CVPR, 2011.'
- p. 4 Section 3.1
Zeiler and Fergus present a visually attrac-tive deconvolutional approach in [42]
- p. 19 Reference [42]
M. Zeiler, G. Taylor, and R. Fergus. Adaptive deconvolu-tional networks for mid and high level feature learning. In CVPR, 2011.
- p. 19 Reference [42]
M. Zeiler, G. Taylor, and R. Fergus. Adaptive deconvolu-tional networks for mid and high level feature learning
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: that a separate two-author Zeiler & Fergus visualization paper may have been intended. On the page, the text names two authors while reference [42] lists three (Zeiler, Taylor, Fergus) and is a feature-learning paper. Its back-reference (page 4) confirms this is the cited entry. Only the authors can say whether the citation or the author naming is what was intended.
Second check: another model agreed
Report a problem with this finding
Wording Figure 6 caption: 'fine-tuning improves robustness for all characteristics including occlusion', but the occlusion lowest subset is 0.212 (R-CNN fc6) and 0.179 (R-CNN FT fc7).
- p. 8 Figure 6 caption and plots (occ)
Instead, fine-tuning improves robustness for all characteristics including occlusion, truncation, viewpoint, and part visibility.
Only the authors can say what is right here.
Judge's note: The page 8 image shows the occlusion values: fc6 has max 0.612 and min 0.212; FT fc7 has max 0.701 and min 0.179. The occlusion spread widens and its lowest subset falls. The caption itself says the improvement holds for 'nearly all characteristics' and that fine-tuning 'does not reduce sensitivity', so the closing 'all ... including occlusion' overstates what the plots show.
Second check: another model agreed
Report a problem with this finding
Typo Introduction, p.2: 'The second principle contribution of this paper' should read 'principal contribution'.
- p. 2 Section 1, right column
The second principle con-
- p. 2 Introduction, right column
The second principle con-tribution of this paper
Judge's note: The page image shows 'The second principle con-tribution of this paper is to show that supervised pre-training...'. It is a wrong-word typo: 'principal' (main) is meant.
Second check: none (judge only)
Report a problem with this finding
Typo Reference [1] back-reference reads 'TPAMI, 2012. 2', but the only citation, 'Examples include: objectness [1]', is on page 3.
- p. 14 Reference list, entry [1]
ness of image windows. TPAMI, 2012. 2
- p. 3 Section 2.1, first line of left column text
Examples include: objectness [1]
Judge's note: The text has only one citation of [1], at the top of page 3, left column. The co-cited [14] and [3] give back-reference 3, so the '2' is likely a backref artefact at the page break. The pointer is wrong but trivial.
Second check: none (judge only)
Report a problem with this finding
Typo Figure 3 caption: 'Methods preceeded by * use outside training data' should read 'preceded'.
- p. 5 Figure 3 caption
Methods preceeded by * use outside training data
Judge's note: The misspelling is visible in the caption on the page 5 image.
Second check: none (judge only)
Report a problem with this finding
Typo Table 1 and Table 2 captions refer to 'Section C' for bounding-box regression, while the body text says 'Appendix C' (and 'Appendix A/B/D/E').
- p. 5 Table 1 caption
Bounding-box regression (BB) is described in Section C.
- p. 6 Table 2 caption
reduces localization errors (Section C)
Judge's note: Both captions read 'Section C' (pp.5-6). The text uses 'Full details are given in Appendix C' and 'Appendix' for the others. The target is unambiguous, so this is only an inconsistency in wording.
Second check: none (judge only)
Report a problem with this finding
Typo Sec. 4.5: 'which is a smaller relative gain that what was observed in PASCAL' should read 'than what'.
- p. 9 Section 4.5, last sentence
which is a smaller relative gain that what was ob-
- p. 9 Section 4.5, end
which is a smaller relative gain that what was ob-served in PASCAL
- p. 9 Section 4.5
smaller relative gain that what was ob-
Judge's note: The text is confirmed in the page 9 image.
Second check: none (judge only)
Report a problem with this finding
Typo Sec. 4.3: 'CNN fine-tuning was run for 50k SGD iteration on val1+trainN' should read 'iterations'.
- p. 9 Section 4.3
for 50k SGD iteration on val1+trainN
Judge's note: The text is confirmed in the page 9 image.
Second check: none (judge only)
Report a problem with this finding
Typo Conclusion: 'for a auxiliary task with abundant data' should read 'an auxiliary task'.
- p. 11 Section 6 Conclusion
for a auxiliary task with abundant data
Judge's note: The text is confirmed in the page 11 image.
Second check: none (judge only)
Report a problem with this finding
Typo Appendix A: 'subtracted before inputing the image into the CNN' should read 'inputting'.
- p. 11 Appendix A
before inputing
Judge's note: The text is confirmed in the page 11 image.
Second check: none (judge only)
Report a problem with this finding
Typo Appendix C: 'Hence forth, we drop the superscript i unless it is needed.' should read 'Henceforth'.
- p. 12 Appendix C, second paragraph
Hence forth, we drop the superscript i unless it is needed.
- p. 12 Appendix C
Hence forth, we drop the superscript
Judge's note: The page 12 image shows 'Hence forth' with a space at the start of a line, not split by hyphenation.
Second check: none (judge only)
Report a problem with this finding
Typo Reference [37] misspells the institution name: "Technical Report A.I. Memo No. 1521, Massachussets Institute of Technology, 1994" (should be "Massachusetts").
- p. 19 Reference [37]
Massachussets Institute of Technology
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: the institution's name is spelled "Massachusetts Institute of Technology". The crop of PDF page 19 (reference [37], Sung and Poggio) shows "Massachussets" as printed, with a doubled 's' in the wrong place. This is a cosmetic misspelling in the reference list, so it is rated S4. It does not affect what the reference points to.
Second check: none (judge only)