Distinctive Image Features from Scale-Invariant Keypoints
International Journal of Computer Vision, 2004 · paper · AI and machine learning
Presents a method for extracting image features invariant to scale and rotation, and robust to viewpoint, noise and lighting changes. It also describes using them for object recognition by matching against a database, which identifies objects despite clutter and occlusion in near real time.
Checked: PDF from cs.ubc.ca, fetched 4 Oct 2026. Page numbers are those of that PDF.
1 finding
Reference or label Wrong section pointer on p9: "using the nearest-neighbor matching procedure to be described in Section 6" — the procedure is described in Section 7 (7.1 Keypoint matching, 7.2 Efficient nearest neighbor indexing).
- p. 9 Section 3.2, paragraph after Figure 4
using the nearest-neighbor matching procedure to be described in Section 6
Judge's note: Page 9 image confirms the pointer text. Section 6 is the local image descriptor (pages 15-17 viewed: 6.1 representation, 6.2 testing, 6.3 affine change). In the text layer, the nearest-neighbour definition ("The best candidate match for each keypoint is found by identifying its nearest neighbor...") sits under Section 7.1, which itself points back to Section 6 only for the descriptor. Section 6.2 also defers the nearest-neighbour methods to "described below". I located Section 7.1 by text search and did not open its page image.
Second check: another model agreed
Report a problem with this finding
11 minor (typos and the like; most checked by the judge only)
Wording Text on p9 says "The lower line on this graph shows the number of keypoints that are correctly matched to a database of 40,000 keypoints", but the first graph of Figure 3 has y-axis "Repeatability (%)" and its caption says the lower line "shows the percent of keypoints".
- p. 9 Section 3.2, first paragraph under Figure 4
The lower line on this graph shows the number of keypoints that are correctly matched to a database of 40,000 keypoints
- p. 8 Figure 3 caption and y-axis of first graph
The lower line shows the percent of keypoints that have their descriptors correctly matched
- p. 8 Figure 3 caption and first-graph y-axis
The lower line shows the percent of keypoints that have their descriptors correctly matched to a large database
- p. 9 Section 3.2
The lower line on this graph shows the number of keypoints that are correctly
Judge's note: Checked the Figure 3 graph and caption on page 8 and the body text on page 9. The first graph plots percentages; counts are in the second graph. "Number" is loose wording for proportion and the meaning is recoverable from the figure, so this is a minor wording inconsistency.
Second check: another model agreed
Report a problem with this finding
Wording Section 5: "The measured variance of orientation for the correct matches is about 2.5 degrees, rising to 3.9 degrees for 10% noise" — a variance is quoted in degrees, not degrees squared.
- p. 13 Section 5, last paragraph
The measured variance of orientation for the correct matches is about 2.5 degrees, rising to 3.9 degrees for 10% noise
Judge's note: Page 13 image confirms the wording. The unit suggests a standard deviation or RMS error is meant; "variance" may be informal. The paper does not define the quantity elsewhere (text search for "variance" found no other use). Minor ambiguity.
Second check: another model agreed
Report a problem with this finding
Typo Abstract has a duplicated word: "robust matching across a a substantial range of affine distortion".
- p. 1 Abstract, line 4
robust matching across a a substantial range
- p. 1 Abstract, sentence 2
robust matching across a a substantial range of
- p. 1 Abstract
across a a substantial range
- p. 1
robust matching across a a substantial range of affine dis-
Judge's note: Page 1 image shows "across a a substantial range" on a single line of the abstract, so it is not a line-break or extraction artefact.
Second check: none (judge only)
Report a problem with this finding
Typo Missing article in Section 2: "between the two views of rigid scene" (should read "of a rigid scene").
- p. 3 Section 2, paragraph 2
between the two views of rigid scene
Judge's note: Page 3 image shows the text exactly as quoted, with no article before "rigid scene". Not an extraction artefact.
Second check: none (judge only)
Report a problem with this finding
Typo Duplicated word in Section 2: "The affine frames are are also more sensitive to noise".
- p. 4 Section 2, paragraph 2
The affine frames are are also more sensitive to noise
- p. 4 Section 2
The affine frames are are also more sensitive
- p. 4
The affine frames are are also more sensitive to noise than
- p. 4
affine space. The affine frames are are also more sensitive
Judge's note: Page 4 image shows "are are" mid-line in the affine-invariance paragraph.
Second check: code check
Report a problem with this finding
Typo Section 3.2: "scaling by a random amount between 0.2 of 0.9 times the original size" — "of" should be "and".
- p. 8 Section 3.2, last paragraph
a random amount between 0.2 of 0.9 times the original size
- p. 8 Section 3.2
would be be present in the matched image
Judge's note: Page 8 image shows "between 0.2 of 0.9 times the original size". "Between" requires a range, so the evident fix is "0.2 and 0.9". The "be be" duplication also mentioned in this cluster is handled under C007.
Second check: none (judge only)
Report a problem with this finding
Typo Duplicated word in Section 3.2: "keypoints would be be present in the matched image".
- p. 8 Section 3.2, last paragraph
keypoints would be be present in the matched image
- p. 8 Section 3.2
would be be present in the matched image
Judge's note: Page 8 image shows "would be be present" on one line in the last paragraph.
Second check: none (judge only)
Report a problem with this finding
Typo Figure 5 caption: "(b) The initial 832 keypoints locations at maxima and minima" — should be "keypoint locations".
- p. 11 Figure 5 caption, part (b)
The initial 832 keypoints locations at maxima and minima
Judge's note: Page 11 image shows "832 keypoints locations" in the Figure 5 caption.
Second check: none (judge only)
Report a problem with this finding
Typo Section 6.1: "It is important to avoid all boundary affects" — should be "boundary effects".
- p. 15 Section 6.1, last paragraph
It is important to avoid all boundary affects
- p. 21 Section 7.3, paragraph 2
To avoid the problem of boundary effects in bin assignment
- p. 15 Section 6.1
avoid all boundary affects
- p. 15 Section 6.1
avoid all boundary affects in which the descriptor abruptly changes
Judge's note: Page 15 image shows "boundary affects". The text layer shows the paper uses "boundary effects" for the same idea in Section 7.3.
Second check: none (judge only)
Report a problem with this finding
Typo Section 9: "to recognizing particular objects categories" should read "particular object categories".
- p. 26 Section 9, second paragraph on the page
to recognizing particular objects categories
- p. 26 Section 9, last paragraph
recognizing particular objects categories
Judge's note: Viewed the crop of page 26: the text reads "suited to recognizing particular objects categories." The following sentences use "object classes". Stray plural; cosmetic.
Second check: none (judge only)
Report a problem with this finding
Typo Reference list, Schaffalitzky and Zisserman 2002: "‘How do I organize my holiday snaps?”’" opens with a single quote but closes with a double quote followed by a single quote.
- p. 28 References, Schaffalitzky and Zisserman 2002
‘How do I organize my holiday snaps?”’
Judge's note: Viewed the crop of page 28: the title opens with a single left quote before "How" and closes after "snaps?" with three closing strokes (a double plus a single closing quote). The marks are unbalanced on the page itself, not an extraction artefact. Cosmetic.
Second check: none (judge only)