Segment Anything
2023 · paper · AI and machine learning
Introduces the Segment Anything project: a promptable segmentation task, a model (SAM), and a dataset of over 1 billion masks on 11 million images. The authors report that the model's zero-shot performance on many tasks is often competitive with prior supervised results.
Checked: PDF from arxiv.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
4 findings
Substance The Fig. 18(f) OVIS legend gives "7.7 ± 0.12, SAM - single output" and "7.2 ± 0.13, SAM", so SAM is rated below its single-output ablation. This conflicts with Table 8 (OVIS SAM > SAM single out.: 3e-10, CI (0.27, 0.63)), with the SAM > RITM CI (1.38, 1.84) beside RITM 6.1, and with the text's claim that single-output ratings are consistently lower. The two values look swapped.
- p. 24 Figure 18(f) OVIS legend
7.7 ± 0.12, SAM - single output
- p. 24 Figure 18(f) OVIS legend
7.2 ± 0.13, SAM
- p. 24 Table 8, OVIS row
(0.27, 0.63)
- p. 9 §7.1 Results
SAM with a single output mask has consistently lower rat-
Only the authors can say what is right here.
Judge's note: On the p24 image, the panel (f) legend reads RITM 6.1, single output 7.7, SAM 7.2, GT 8.8. In Table 8 the OVIS row gives CI99 (1.38, 1.84) for SAM vs RITM and (0.27, 0.63) for SAM vs single output, and the caption says all CIs exclude zero. As printed, the legend gives a SAM minus RITM difference of 1.1 and a SAM minus single-output difference of -0.5, both outside their CIs. Swapping the two values gives 1.6 and 0.5, both inside. In every other panel SAM is above single output. The bars in panel (f) also show SAM higher than single output at ratings 9 and 10. Only the authors can say which values are correct.
Second check: another model agreed
Report a problem with this finding
Substance §7.3 says SAM 'outperforms ViTDet-H on medium and large objects' and 'only underperforms ViTDet-H on small objects and frequent objects', but in Table 4 the large column reads ViTDet-H 87.0 and SAM 86.9.
- p. 10 §7.3 Results
it outperforms ViTDet-H on medium and large objects, as well as rare and common objects
- p. 10 Table 4, 'large' column
87.0
- p. 10 Table 4, SAM row
86.9
- p. 10 Table 4, large column
87.0 ... 86.9
Judge's note: Checked Table 4 on the p.10 image: large column, ViTDet-H 87.0 and SAM 86.9. SAM is 0.1 lower on large objects, so 'outperforms ... on large objects' and the 'only underperforms on small and frequent' list conflict with the printed table. Medium (81.6 vs 80.8), rare and common do support the claim. The gap is small, but the printed direction is the opposite of the text.
Second check: another model agreed
Report a problem with this finding
Reference or label LVIS v0.5 is cited as [17] in §E ("We choose LVIS v0.5 [17], ... IBD [17]") and in the Fig. 18(a) caption ("(a) LVIS v0.5 [17]"). [17] is the MVS buildings (IBD) paper; LVIS is [44], as Table 8 gives it.
- p. 23 §E Datasets
We choose LVIS v0.5 [17]
- p. 24 Figure 18(a) caption
(a) LVIS v0.5 [17]
- p. 13 Reference [17]
3D instance segmentation of MVS buildings
- p. 24 Table 8
LVIS v0.5 [44]
Judge's note: On p23 the §E Datasets sentence cites both LVIS v0.5 and IBD as [17]. On p24 the Fig. 18(a) caption reads 'LVIS v0.5 [17]', while Table 8 on the same page reads 'LVIS v0.5 [44]'. On p14, ref [44] is Gupta et al., 'LVIS: A dataset for large vocabulary instance segmentation'. Per the cluster's crop of p13, [17] is '3D instance segmentation of MVS buildings'.
Second check: another model agreed
Report a problem with this finding
Reference or label Table 7 TrashCan row: "Images are sourced from the J-EDI [69] dataset." But [69] is Lucas et al., "Gelatinous zoo-plankton biomass in the global oceans: geographic variation and environmental drivers", Global Ecology and Biogeography, 2014.
- p. 20 Table 7, TrashCan row
Images are sourced from the J-EDI [69] dataset.
- p. 14 Reference [69]
Gelatinous zoo-plankton biomass in the global oceans: geographic variation and environmental drivers
Only the authors can say what is right here.
Judge's note: Rests on outside knowledge: J-EDI (the source of the TrashCan images) is JAMSTEC's E-library of Deep-sea Images, not the Jellyfish Database Initiative (JeDI) behind the Lucas et al. paper. I checked both pages. Table 7 on p20 cites [69] for J-EDI, and the [69] entry on p14 is the zooplankton biomass ecology paper, which by its title is not an underwater image dataset. Only the authors can say what the intended reference was.
Second check: another model agreed
Report a problem with this finding
18 minor (typos and the like; most checked by the judge only)
Wording The introduction says without qualification 'we found that SAM performs similarly across different groups of people'. Yet the Table 2 caption says 'all confidence intervals overlap except older vs. middle', and §6/§C report 'an indication of bias across perceived gender presentation' for clothing.
- p. 2 §1 Introduction, 'Responsible AI (§6)' paragraph
we found that SAM performs sim-ilarly across different groups of people
- p. 8 Table 2 caption
all confidence intervals overlap except older vs. middle.
- p. 8 §6 Fairness in segmenting people, last sentence
we find an indication of bias across perceived gender presentation.
Judge's note: Checked the intro on p.2 and Table 2 with its caption and §6 on p.8. The perceived-age intervals for older and middle are disjoint at both 1 and 3 points, and the paper itself says so. §6 also points to a clothing bias by gender presentation (Table 6). The intro's summary leaves out these exceptions that the paper reports. This is a wording point; the data are fine.
Second check: another model agreed
Report a problem with this finding
Wording §7.1 says of Fig. 9b 'all differences are significant; see §E for details'. But §E/Table 8 test only SAM > baseline and SAM > SAM single-output, while Fig. 9b also plots ground truth (e.g. IBD: '8.3 ± 0.08, SAM' vs '8.4 ± 0.09, GT').
- p. 9 §7.1 Results
(all differences are significant; see §E for details)
- p. 24 Figure 18(d) IBD legend
8.3 ± 0.08, SAM
- p. 24 Figure 18(d) IBD legend
8.4 ± 0.09, GT
Only the authors can say what is right here.
Judge's note: Page 24: §E says statistical tests were run 'for two hypotheses', (1) SAM vs baseline and (2) SAM vs single-output SAM, and Table 8 lists only those. No test involving the ground-truth ratings plotted in Fig. 9b is reported. 'All differences' therefore claims more than the reported tests cover. Query for the authors: does 'all differences' mean only the tested SAM comparisons? I make no claim here about whether any untested difference is significant.
Second check: another model agreed
Report a problem with this finding
Wording §C says 'The gap closes when moving from 1 to 3 point evaluation.', yet Table 6 at 3 points still shows feminine '90.7 ±0.5' vs masculine '92.3 ±0.4'.
- p. 19 §C Fairness in segmenting clothing
The gap closes when moving from 1 to 3 point evaluation.
- p. 19 Table 6, 3 points
90.7 ±0.5
- p. 19 Table 6, 3 points
92.3 ±0.4
Judge's note: Checked Table 6 on p.19. The gap narrows from 4.7 to 1.6, and the printed 3-point intervals still do not overlap. 'Closes' most likely means 'narrows'. The text's following statement, which limits the bias to the one-point prompt, adds to the ambiguity. Wording only.
Second check: another model agreed
Report a problem with this finding
Typo Figure 6 (left) x-axis bins are labelled '<10', '11-50', '51-100', '101-200', '>200', so as printed a count of exactly 10 masks falls in no bin.
- p. 7 Figure 6 left, x-axis
<10
Judge's note: Read the bin labels on the p.7 page image. As printed they do not cover the value 10. Almost certainly a labelling slip ('≤10' or '10-50' intended); the data are not affected.
Second check: none (judge only)
Report a problem with this finding
Typo Table 3 lists the method as 'EDETR [79]', while reference [79] is titled 'EDTER: Edge detection with transformer'.
- p. 10 Table 3
EDETR [79]
- p. 15 Reference [79]
EDTER: Edge detection with transformer
- p. 15 Reference [79]
EDTER: Edge detection with transformer.
Judge's note: The p.10 page image shows 'EDETR' in Table 3, and the crop of reference [79] on p.15 reads 'EDTER'. The method name is misspelled in the table.
Second check: none (judge only)
Report a problem with this finding
Typo §7.4 reads 'we run a object detector', which should be 'an object detector'.
- p. 10 §7.4 Approach
we run a object detector
Judge's note: Confirmed in the §7.4 Approach paragraph on the p.10 page image.
Second check: none (judge only)
Report a problem with this finding
Typo Reference [90] title reads 'mulit-class object recognition and segmentation', which should be 'multi-class'.
- p. 15 Reference [90]
mulit-class object recognition and segmentation
Judge's note: Seen in the reference-list crop on p.15.
Second check: none (judge only)
Report a problem with this finding
Typo §A reads 'For ablations and others variations on training', which should be 'other variations'.
- p. 18 §A, training recipe deviations
For ablations and others variations on training
Judge's note: Seen in the p.18 crop.
Second check: none (judge only)
Report a problem with this finding
Typo §D.3 Baseline: "we disable score thresholding in the three cascade stages and as raise the maximum number of predictions per stage to 1000" has a stray 'as'.
- p. 21 §D.3 Baseline
stages and as raise the
- p. 21 D.3 Baseline
and as raise the
Judge's note: The full page image of p21 (right column, Baseline paragraph) shows 'and as raise the'. The 'as' is an extra word.
Second check: none (judge only)
Report a problem with this finding
Typo §D.3: "we compare to an ab-lated version SAM that only outputs a single mask" is missing 'of' (should read 'ablated version of SAM').
- p. 21 §D.3 Method
an ab-lated version SAM that only outputs a single mask
Judge's note: The p21 image (right column, second paragraph of Method) reads 'an ablated version SAM that only outputs a single mask instead of three'. The word 'of' is missing.
Second check: none (judge only)
Report a problem with this finding
Typo §D.3 Method: "the input point grid and the NMS threshold duplicate mask suppression" is missing 'for'.
- p. 21 D.3 Method
the NMS threshold duplicate mask suppression
Judge's note: The p21 image (right column, Method paragraph) shows 'the NMS threshold duplicate mask suppression'. A preposition is missing.
Second check: none (judge only)
Report a problem with this finding
Typo §D.6: "capture any semantics in its representation even though is not trained with explicit semantic supervision" is missing 'it'.
- p. 22 §D.6
even though is not trained
Judge's note: The crop from p22, §D.6, reads 'even though is not trained'. The subject 'it' is missing.
Second check: none (judge only)
Report a problem with this finding
Typo "asked to judge the mask on three criterion" (§E) and "Judge the quality of the mask on three criterion" / "according to the three criterion" (Fig. 19 guidelines) use the singular where the plural 'criteria' is needed.
- p. 23 §E Methodology
mask on three criterion
- p. 29 Figure 19, Task panel
Judge the quality of the mask on three criterion
Judge's note: Seen on the p23 image in the right column, Methodology paragraph, and in the Fig. 19 Task panel crop from p29.
Second check: none (judge only)
Report a problem with this finding
Typo §E: "a low score (2-4) indicates that the mask has huge errors, such including huge regions of other objects" is garbled (should read 'such as including' or 'including').
- p. 23 §E Methodology
such including huge regions of other objects
- p. 23 E Methodology
errors, such including huge regions
Judge's note: The p23 image, right column, second paragraph, reads 'errors, such including huge regions of other objects'.
Second check: none (judge only)
Report a problem with this finding
Typo Data Annotation Card (p.27) has two grammar slips: 'Depending on annotators’s skills' (should be annotators') and 'one of the most geographically diverse segmentation dataset' (should be datasets).
- p. 27 §F.2 Task Formulation Q1
Depending on annotators’s skills
- p. 27 §F.2 Selecting Annotations Q5
one of the most geographically diverse segmentation dataset
Judge's note: Both crops of page 27 show the text exactly as quoted. These are copyedit errors only.
Second check: none (judge only)
Report a problem with this finding
Typo Model card (Table 9, p.28) has the field label 'Mode type'; it should read 'Model type'.
- p. 28 Table 9, Model Overview
Mode type
Judge's note: The crop shows 'Mode type | Promptable segmentation model' in the Model Overview. This is a missing-letter typo.
Second check: none (judge only)
Report a problem with this finding
Typo 'complete guidelines given to annotations' (Fig. 19 and 20 captions and §G) should say annotators, and 'Some images been edited slightly' (Fig. 19 and 20 captions) is missing 'have'.
- p. 29 Figure 19 caption
complete guidelines given to annotations for the human review of mask quality. Some images been edited slightly
- p. 29 §G
complete guidelines given to annotations
- p. 30 Figure 20 caption
Some images been edited slightly
Judge's note: The page 29 crop shows the Fig. 19 caption and the §G text, and the page 30 crop shows the Fig. 20 caption, all worded exactly as quoted. These are copyedit errors only.
Second check: none (judge only)
Report a problem with this finding
Typo Fig. 19 example caption has a duplicated word: 'In top top image, the point is on one orange rind'.
- p. 29 Figure 19, 'Include an arbitrary part of a collection' example
In top top image
- p. 29
ary part of a collection’: In top top image, the point is
Judge's note: The crop of page 29 shows 'In top top image' in the 'Include an arbitrary part of a collection' example. The word 'top' is duplicated.
Second check: code check