Fast R-CNN
2015 · paper · AI and machine learning
Proposes Fast R-CNN, a single-stage training method for object detection that jointly classifies proposals and refines their locations. It trains a VGG16 network 9× faster than R-CNN, tests 213× faster, and reaches higher accuracy on PASCAL VOC 2012.
Checked: PDF from arxiv.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
0 findings
None that passed a second check.
4 minor (typos and the like; most checked by the judge only)
Reference or label "The difference in mAP when learning from conv2 1 up was only +0.3 points (Table 5, last column)": the last column of Table 5 is SPPnet L ≥ fc6 (63.1), while the conv2_1 result (67.2) is in the third column.
- p. 6 Section 4.5, text
was only +0.3 points (Ta-ble 5, last column)
- p. 6 Table 5
SPPnet L
- p. 6 Section 4.5, second paragraph
The difference in mAP when learning from conv2 1 up was only +0.3 points (Table 5, last column)
Judge's note: Table 5 on the page 6 image has four columns: ≥fc6 61.4, ≥conv3_1 66.9, ≥conv2_1 67.2 (all under the 'model L' heading), then SPPnet L ≥fc6 63.1. Since 67.2−66.9=0.3, the text refers to the ≥conv2_1 column. That column is the last one in the 'model L' group, not the last column of the table, so the pointer is ambiguous. It should name the ≥ conv2_1 column.
Second check: another model agreed
Report a problem with this finding
Wording SPPnet VGG16 training time is "25.5 hours" in Section 4.4 (giving "2.7×") but "25" h in Table 4, whose 3.4× train speedup over R-CNN's 84 h matches 25 h.
- p. 6 Section 4.4 text
2.7× faster (in 9.5 vs. 25.5 hours)
- p. 6 Table 4, SPPnet †L, train time (h)
25
- p. 6 Section 4.4, first paragraph
trains VGG16 2.7× faster (in 9.5 vs. 25.5 hours)
- p. 1 Abstract / Introduction
Compared to SPPnet, Fast R-CNN trains VGG16 3× faster
Only the authors can say what is right here.
Judge's note: Page 6 image: the text says "2.7× faster (in 9.5 vs. 25.5 hours)". Table 4, SPPnet †L column: train time 25, train speedup 3.4×. The R-CNN times in the table are whole hours (22, 28, 84), so 25 could just be 25.5 shown as a whole number. However, 84/25.5 gives 3.3×, not the printed 3.4×, so the table and text do not use one consistent value. The gap is small, and the abstract's "3×" is a rounded summary of either value. The authors should say which value is right.
Second check: another model agreed
Report a problem with this finding
Wording Table 1, tv column: R-CNN BB's "71.1" is bold, but SPPnet BB's "73.5" in the same column is higher and is not bold.
- p. 5 Table 1, tv column
71.1
- p. 5 Table 1, SPPnet BB row, tv column
73.5
Only the authors can say what is right here.
Judge's note: Checked bold weight in the Table 1 crop. In every other column the highest value is bold, and that includes the FRCN 07 \ diff row (79.0, 78.6, 48.0), which uses the same training set as SPPnet. tv is the only column where SPPnet has the highest value, and there the bold goes to R-CNN instead. Neither the caption nor the text says SPPnet is excluded from bolding. The data are not affected; only the marking of the best result is inconsistent, so this goes to the authors as a query.
Second check: another model agreed
Report a problem with this finding
Typo "forth column per group" should read "fourth column per group" (Table 6 caption and end of Section 5.1).
- p. 7 Table 6 caption
Multi-task training (forth column per group)
- p. 7 Section 5.1, end
performs multi-task training (forth column per group)
- p. 7 Section 5.1, last paragraph
(forth column per group)
Judge's note: The page 7 image shows "Multi-task training (forth column per group)" in the Table 6 caption, and the same spelling in the Section 5.1 text. It is a misspelling of the ordinal 'fourth'.
Second check: none (judge only)