SSD: Single Shot MultiBox Detector
2016 · paper · AI and machine learning
Presents SSD, an object detector using a single network that predicts scores and box adjustments for default boxes across multiple feature maps, without proposal generation. It reports 74.3% mAP at 59 FPS on VOC2007 and 76.9% mAP at larger input.
Checked: PDF from arxiv.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
4 findings
Substance Abstract: "for 512 x 512 input, SSD achieves 76.9% mAP", but Table 1 (SSD512, 07+12), Table 6 and Table 7 all give 76.8 for SSD512 on VOC2007 test.
- p. 1 Abstract
for 512 × 512 input, SSD achieves 76.9% mAP
- p. 8 Table 1, SSD512 07+12 row
76.8 82.4 84.7 78.4
- p. 15 Table 7, SSD512 rows
76.8
- p. 1 Abstract
SSD achieves 76.9% mAP
Judge's note: Page 1 abstract reads 76.9%. Table 1 (p8) SSD512 07+12 mAP is 76.8; Table 6 (p13) SSD512 VOC2007 07+12 is 76.8; Table 7 (p15) gives 76.8 in both SSD512 rows. Sec 3.1 says SSD512 is 3.6% better than Faster R-CNN (73.2), which gives 76.8. Footnote 1 gives 79.8 for the improved-augmentation model, so it is not 76.9 either. No setting in the paper gives 76.9.
Second check: another model agreed
Report a problem with this finding
Substance "SSD512 is 5.3% better in mAP@0.75" than Faster R-CNN, but Table 5 gives 27.8 vs 23.5, a difference of 4.3.
- p. 11 Sec. 3.4 COCO, text
SSD512 is 5.3% better in mAP@0.75
- p. 12 Table 5, 0.75 column
27.8
- p. 12 Table 5
SSD512 trainval35k 26.8 46.5 27.8
- p. 12 Table 5
46.5 27.8
Judge's note: Page 12 Table 5: SSD512 mAP@0.75 is 27.8, Faster [25] is 23.5 and ION is 23.6. The gain at 0.5 (46.5-45.3=1.2) matches the text. No comparator in Table 5 gives 5.3.
Second check: another model agreed
Report a problem with this finding
Reference or label Reference 15 (Simonyan & Zisserman, "Very deep convolutional networks for large-scale image recognition") gives the venue as "In: NIPS. (2015)".
- p. 17 Reference 15
Very deep convolutional networks for large-scale image recog-nition. In: NIPS. (2015)
- p. 17 Reference 15
nition. In: NIPS. (2015)
- p. 17 Ref 15
very deep convolutional networks for large-scale image recog-
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: the VGG paper appeared at ICLR 2015, not NIPS. The page 17 image confirms the entry reads NIPS (2015).
Second check: another model agreed
Report a problem with this finding
Wording Text says accuracy drops "monotonically from 74.3 to 62.4", but Table 3's 'Yes' column reads 74.3, 74.6 (bold), 73.8, 70.7, 64.2, 62.4. It rises at the second row.
- p. 10 Multiple output layers paragraph
dropping monotonically from 74.3 to 62.4
- p. 10 Table 3, 'Yes' column
74.6
- p. 10 Table 3 row 2
74.6 63.1 8764
Judge's note: The page 10 image confirms the Table 3 values. The quoted endpoints identify the 'Yes' column, and that column is not monotone. The overall decrease holds; the word 'monotonically' does not.
Second check: another model agreed
Report a problem with this finding
8 minor (typos and the like; most checked by the judge only)
Wording Eq. (2) writes the indicator as x^k_ij, but it is defined as x^p_ij ("Let x^p_ij = {1, 0} be an indicator ... of category p"), and k is not defined at this point.
- p. 5 Eq. (2), localization loss
x^k_ij smoothL1(l^m_i − ĝ^m_j)
- p. 5 Training objective definition
Let x^p_ij = {1, 0} be an indicator
Judge's note: The page 5 image confirms that Eq. (2) uses superscript k, while the definition and Eq. (3) use p. Sec 2.2 later uses k for the feature-map index. This is a notation inconsistency; the evident fix is x^p_ij.
Second check: another model agreed
Report a problem with this finding
Wording "If we only use conv7 for prediction, the performance is the worst" holds for Table 3's 'Yes' column (62.4) but not for its 'No' column (64.0 vs 63.1 and 63.4), and the sentence follows the discussion of pruning boundary boxes.
- p. 10 Sec. 3.2, 'Multiple output layers' paragraph
If we only use conv7 for prediction, the performance is the worst
- p. 10 Table 3, last row, 'No' column
64.0
Only the authors can say what is right here.
Judge's note: Page 10: the surrounding sentences ('after the pruning', 'performance starts increasing again') describe the 'No' column, where conv7-only (64.0) is not the minimum. The text does not say which column the sentence refers to.
Second check: another model agreed
Report a problem with this finding
Wording "We have seen a consistent increase of 2%-3% mAP across multiple datasets, as shown in Table 6", but the Table 6 gains range from 1.6 to 3.6.
- p. 12 Sec. 3.6
We have seen a consistent increase of 2%-3% mAP across multiple datasets, as shown in Table 6
- p. 13 Table 6
81.2
Judge's note: Page 13 Table 6 gains: VOC07 +2.9/+3.0; 07+12+COCO +1.6/+1.6; VOC12 +3.4/+3.6; VOC12+COCO +1.8/+2.2; COCO +1.9 to +2.5. The stated range is approximate wording; several gains fall outside it.
Second check: another model agreed
Report a problem with this finding
Typo Duplicated word: "bounding box hypotheses and and is as accurate as approaches that do".
- p. 2 Introduction, 2nd paragraph
hypotheses and and is as accurate
- p. 2 Introduction, para 2
bounding box hypotheses and and is as accurate as ap-
- p. 2
resample pixels or features for bounding box hypotheses and and is as accurate
- p. 2
s for bounding box hypotheses and and is as accurate as a
Judge's note: The page 2 image shows an italic 'and' followed immediately by a roman 'and' on the same line. Only one is needed.
Second check: code check
Report a problem with this finding
Typo Typo: "If N = 0, wet set the loss to 0." should read "we set".
- p. 5 after Eq. (1)
If N = 0, wet set the loss to 0.
Judge's note: The page 5 image shows 'wet set'.
Second check: none (judge only)
Report a problem with this finding
Typo Grammar: "the default boxes do not necessary need to correspond" should be "necessarily".
- p. 6 Choosing scales paragraph
the de-fault boxes do not necessary need to correspond
- p. 6 Sec 2.2, Choosing scales
default boxes do not necessary need to correspond
Judge's note: Seen in the page 6 image, in the paragraph before Eq. (4).
Second check: none (judge only)
Report a problem with this finding
Typo Grammar: "Both our SSD300 and SSD512 method outperforms Faster R-CNN" should be "methods outperform".
- p. 14 Sec 3.7
Both our SSD300 and SSD512 method outperforms Faster R-CNN
Judge's note: Seen in the page 14 image.
Second check: none (judge only)
Report a problem with this finding
Typo Reference 12: "ParseNet: Looking wider to see better. In: ILCR. (2016)". The venue is misspelled; elsewhere it is ICLR.
- p. 17 Reference 12
ParseNet: Looking wider to see better. In: ILCR. (2016)
- p. 17 Reference 12
In: ILCR. (2016)
Judge's note: Seen in the page 17 image. References 13 and 17 spell it ICLR.
Second check: none (judge only)