End-to-End Object Detection with Transformers
2020 · paper · AI and machine learning
Presents DETR, which treats object detection as direct set prediction using a transformer encoder-decoder and a bipartite matching loss, removing hand-designed parts like anchors and non-maximum suppression. It matches Faster R-CNN on COCO and extends to panoptic segmentation.
Checked: PDF from arxiv.org, fetched 4 Oct 2026. Page numbers are those of that PDF.
11 findings
Substance The baseline DETR (6 encoder layers, 86 GFLOPS, 41.3M parameters) is given 23 FPS in Table 2 ("86/23") but 28 FPS in Table 1 ("86/28") and in Section 4.2 ("and runs at 28 FPS").
- p. 10 Table 2, row '6' layers, GFLOPS/FPS
86/23
- p. 9 Table 1, row DETR, GFLOPS/FPS
86/28
- p. 10 Section 4.2, first paragraph
and runs at 28 FPS
- p. 10 Section 4.2, first paragraph
and runs at 28 FPS, similarly to Faster R-CNN-FPN with the same backbone
Only the authors can say what is right here.
Judge's note: Read Table 1 on page 9 (DETR 86/28, 41M) and Table 2 on page 10 (6 layers: 86/23, 41.3M, AP 40.6; 0 layers: 76/28). Section 4.2 on the same page describes the 6-encoder, 6-decoder, 41.3M model with 40.6 AP as running at 28 FPS. Same architecture and same GFLOPS, so the two FPS figures are for one quantity. No statement of different timing conditions found in the text or in A.3/A.4. Only the authors can say which figure is right. The part of this cluster about APL +7.8 is the separate defect judged under C019.
Second check: another model agreed
Report a problem with this finding
Substance Section 4.1 says DETR is "improving APL (+7.8), however note that the model is still lagging behind in APS (-5.5)", but Table 1 gives +7.7 APL (61.1 vs 53.4) and -6.1 APS (20.5 vs 26.6) against Faster RCNN-FPN+, and no other Faster R-CNN row gives the quoted pair.
- p. 10 Section 4.1, paragraph after Table 2
improving APL (+7.8), however note that the model is still lagging behind in APS (-5.5)
- p. 9 Table 1, row DETR
42.0 62.4 44.2 20.5 45.8 61.1
- p. 9 Table 1, row Faster RCNN-FPN+
42.0 62.1 45.5 26.6 45.4 53.4
- p. 9 Table 1
DETR 86/28 41M 42.0 62.4 44.2 20.5 45.8 61.1
Only the authors can say what is right here.
Judge's note: Read Table 1 on page 9 and the sentence on page 10. The comparison is with Faster R-CNN with the same number of parameters at 42 AP, which is Faster RCNN-FPN+ (42M, 42.0 AP). The APL figure is off by 0.1 and could come from unrounded values, but the APS gap differs by 0.6, which rounding cannot explain. Against the plain FPN row the gaps are +9.1 and -3.7; against DC5+ +6.1 and -2.4. No explanation in the text. Either the text or the table is out of date; the authors must say which.
Second check: another model agreed
Report a problem with this finding
Substance "By reducing the number of network parameters from 41.3M to 28.7M, leaving only 10.8M in the transformer" conflicts with Section 4.1's "23.5M are in ResNet-50, and 17.8M are in the transformer": 28.7M total minus the 23.5M backbone leaves 5.2M, not 10.8M.
- p. 11 Section 4.2, Importance of FFN
from 41.3M to 28.7M, leaving only 10.8M in the transformer
- p. 10 Section 4.1
out of which 23.5M are in ResNet-50, and 17.8M are in the transformer
- p. 10 Sec 4.1
23.5M are in ResNet-50, and 17.8M are in the transformer
- p. 11 Importance of FFN
leaving only 10.8M in the transformer
Only the authors can say what is right here.
Judge's note: Read page 11 (Importance of FFN) and page 10 (Section 4.1). The backbone is unchanged in this ablation. 23.5 + 17.8 = 41.3 matches the stated total, so "transformer" covers everything outside the backbone. After removing the FFNs, 28.7 - 23.5 = 5.2M would remain, while 23.5 + 10.8 = 34.3M would be the total if 10.8M were right. No definition elsewhere in the paper reconciles the figures. One of 28.7M or 10.8M is wrong; only the authors can say which.
Second check: another model agreed
Report a problem with this finding
Substance FFN layers are said to "have Md input and output channels in our case", but by the paper's definitions (d' = d/M, "Md' = d") the input and output of these layers have d = Md' channels, not M times d.
- p. 21 Appendix A.1, Feed-forward network (FFN) layers
input and output channels in our case
- p. 22 Appendix A.3, Computational complexity
embeddings (and M d′ = d)
Only the authors can say what is right here.
Judge's note: Viewed pages 20, 21 and 22. Page 20 defines d' = d/M and gives the attention output as d x N_q; page 22 states 'Md' = d'. The FFN follows the attention output with a residual connection, so its input and output width is d. 'Md' on page 21 clearly prints without a prime (checked in the image, not only the text layer). Most likely a missing prime (Md'), but the authors might have meant the hidden width, so this is a query for them. I did not check the reviewer's 8 x 256 = 2048 figure against the paper.
Second check: another model agreed
Report a problem with this finding
Reference or label "We describe the architecture in more details in the supplement" and "The new mask head is trained for 25 epochs (see supplementary for details)" point to material that the appendix (A.1-A.6) does not contain.
- p. 15 Section 4.4
We describe the architecture in more details in the supplement.
- p. 16 Section 4.4, Training details
The new mask head is trained for 25 epochs (see supplementary for details)
- p. 15 4.4
more details in the supplement
Only the authors can say what is right here.
Judge's note: Both pointers read on the page-15 and page-16 images. Searched the appendix text (PDF pages 20-26): A.1 attention layers, A.2 losses (DICE definition only), A.3 architecture of the detection transformer, A.4 training hyperparameters for detection, A.5 additional results (panoptic qualitative figures), A.6 code. No description of the mask-head architecture beyond Fig. 8 and no mask-head training details. The inconsistent naming ('supplement', 'supplementary', 'supplementary material') is a cosmetic side point.
Second check: another model agreed
Report a problem with this finding
Reference or label A.1 says "except for the details of positional encodings (see Equation 8) that follows [7]", but Equation 8 is the softmax attention-weight formula; the positional encodings Pq and Pkv are added in Equation 7.
- p. 20 Appendix A.1, first paragraph
except for the details of positional encodings (see Equa-
- p. 20 Equation 7
[Q; K; V ] = [T
- p. 20 Appendix A.1, first paragraph
except for the details of positional encodings (see Equation 8) that follows [7]
- p. 20 Equation 7
[Q; K; V ] = [T ′ 1(Xq + Pq); T ′ 2(Xkv + Pkv); T ′ 3Xkv]
Judge's note: Page-20 image: Eq. 7 is [Q;K;V] = [T'1(Xq+Pq); T'2(Xkv+Pkv); T'3 Xkv], introduced with 'after adding the query and key positional encodings [7]'. Eq. 8 defines alpha_{i,j} as a softmax of Q and K dot products and has no positional term. The pointer should be to Equation 7.
Second check: another model agreed
Report a problem with this finding
Wording Section 4.1 reads "a model with 6 transformer and 6 decoder layers of width 256"; "transformer" should be "encoder" (Section 4.2: "model with 6 encoder, 6 decoder layers and width 256").
- p. 10 Section 4.1, paragraph after Table 2
model with 6 transformer and 6 decoder layers of width 256
- p. 10 Section 4.2
model with 6 encoder, 6 decoder layers and width 256
- p. 10 Sec 4.1
a model with 6 transformer and 6 decoder layers
- p. 10 4.1
6 transformer and 6 decoder layers
Judge's note: Read both sentences on page 10. The decoder layers are transformer layers too, so "transformer and decoder" is not a meaningful contrast, and the same model is described as 6 encoder and 6 decoder layers a few paragraphs later and in Table 2. Wrong word with an evident fix.
Second check: another model agreed
Report a problem with this finding
Wording Sec. 4.3 says "In particular, all slots have the mode for predicting image-wide boxes", while the Fig. 7 caption says "almost all slots have a mode of predicting large image-wide boxes".
- p. 14 Fig. 7 caption
We note that almost all slots have a mode of predicting large image-wide boxes
- p. 14 Section 4.3, Decoder output slot analysis
In particular, all slots have
- p. 14 Section 4.3, Decoder output slot analysis
In particular, all slots have the mode for predicting image-wide boxes
- p. 14 Fig. 7 caption
almost all slots have a mode of predicting large image-wide boxes
Only the authors can say what is right here.
Judge's note: Read both on the page-14 image. Text and caption describe the same figure with different quantifiers. In the 20 panels shown, the red central band is clear in some and not evident in several, which fits 'almost all'. No statement elsewhere reconciles the two. Wording inconsistency only.
Second check: another model agreed
Report a problem with this finding
Wording A.1 introduces a double-accented X with superscript (q) as "the concatenation of attention heads", but Equations 5 and 6 call that quantity X'_q; the introduced symbol never appears again.
- p. 20 Appendix A.1, sentence before Equation 5
the concatenation of attention heads, we have
- p. 20 Equation 5
= [attn(Xq, Xkv, T1); ...; attn(Xq, Xkv, TM)]
- p. 20 A.1 Multi-head
¯¯X(q) the concatenation of attention heads
- p. 20 Eq. 5
X′ q = [attn(Xq, Xkv, T1); ...; attn(Xq, Xkv, TM)]
Judge's note: Viewed page 20. The sentence before Eq. 5 uses a doubly accented X^(q); Eq. 5 defines X'_q = [attn(...); ...; attn(...)] and Eq. 6 uses X'_q. The symbol in the sentence is not used anywhere else on the page. Notation inconsistency with an evident fix (use X'_q in the sentence).
Second check: another model agreed
Report a problem with this finding
Wording "B(b_σ(i), b̂_i) means the largest box containing b_σ(i), b̂_i": the enclosing box in the generalized IoU should be the smallest box containing both.
- p. 21 Appendix A.2, Box loss, after Equation 10
largest box containing
Judge's note: Viewed page 21. A 'largest box containing' two boxes is not well defined (any bigger box also contains them), and the same sentence says the areas involving B are computed from min/max of the box coordinates, which yields the smallest enclosing box. The error is evident from the paper's own text; that the cited GIoU definition [38] uses the smallest enclosing box is outside knowledge and only supports it.
Second check: another model agreed
Report a problem with this finding
Wording "the model detects all instances when up to 50 are visible" (and "only 50 instances that are all detected") overstates Fig. 12, where the person and apple curves are already above zero at 50 visible instances.
- p. 24 Appendix A.5, Increasing the number of instances
the model detects all instances when up to 50 are
- p. 25 Fig. 12
% of missed instances
- p. 24 Appendix A.5, Increasing the number of instances
the model detects all instances when up to 50 are visible
Judge's note: Viewed pages 24 and 25. In Fig. 12 at 50 visible instances the person curve is at roughly 5-8% missed, apple at about 4-5% and dog at about 2% (read by eye), and the apple curve is near 10% at the smallest counts. The paragraph's own closing phrase 'near-perfect detections up to 50 objects' matches the figure; 'all' does not.
Second check: another model agreed
Report a problem with this finding
30 minor (typos and the like; most checked by the judge only)
Wording Box-loss index convention differs: Sec. 3.1 writes "L_box(b_i, b̂_σ(i))" (permutation on the prediction) while Appendix A.2 Eq. 9-10 write "L_box(b_σ(i), b̂_i) = λ_iou L_iou(b_σ(i), b̂_i) + λ_L1 ||b_σ(i) − b̂_i||_1" (permutation on the ground truth).
- p. 6 Section 3.1, Bounding box loss
where λiou, λL1 ∈ R are hyperparameters
- p. 21 Appendix A.2, Equation 9
together with a ℓ1 loss on
Judge's note: Viewed pages 6 and 21. Main text (Eq. 1-2 and the box-loss paragraph) consistently attaches σ(i) to the prediction; the appendix, which says it presents the same losses in detail, attaches σ(i) to the ground-truth box and i to the prediction, including in the prose after Eq. 10. Equivalent up to inverting the permutation, so no wrong result, but the notation is inconsistent with the definition of σ in Sec. 3.1.
Second check: another model agreed
Report a problem with this finding
Typo Doubled full stop around the footnote mark: "standard CNN and transformer classes.¹." (PDF p. 2).
- p. 2 Introduction, third paragraph
and transformer classes.
Judge's note: Viewed page 2 image: the sentence ends with a period, then superscript footnote mark 1, then a second period. Real stray punctuation, not an extraction artefact.
Second check: none (judge only)
Report a problem with this finding
Typo Subject-verb agreement: "The design ethos of DETR easily extend to more complex tasks" (p. 3); should be "extends".
- p. 3 Introduction, last paragraph
The design ethos of DETR easily extend
- p. 3 Section 2, first sentence
Our work build on prior work
- p. 2 Introduction
transformer classes.1.
Judge's note: Viewed page 3: 'The design ethos of DETR easily extend' is printed as quoted. This cluster also bundles two items covered elsewhere: 'Our work build on' (C004) and the doubled period after footnote 1 (C001); both are real but counted in those clusters. The item unique to this cluster is 'extend'.
Second check: none (judge only)
Report a problem with this finding
Typo Misspellings: "outperfoms competitive baselines" (p. 3), "FFN inside tranformers" (p. 11), "panoptic segmenation approach" (p. 16); should be outperforms, transformers, segmentation.
- p. 3 Introduction, last paragraph
outperfoms
- p. 3 Introduction
outperfoms competitive baselines
- p. 11 Importance of FFN
FFN inside tranformers
- p. 16 Main results
panoptic segmenation approach
Judge's note: All three misspellings are visible in the page image / crops (p. 3 full page, C002-c for p. 11, C002-d for p. 16). The 'Our work build on' item one reviewer added here is a grammar slip handled under C004.
Second check: none (judge only)
Report a problem with this finding
Typo Subject-verb agreement: "Our work build on prior work in several domains" (p. 3, Sec. 2); should be "builds".
- p. 3 Related work
Our work build on prior work
Judge's note: Viewed page 3; text printed as quoted at the start of Section 2.
Second check: none (judge only)
Report a problem with this finding
Typo Agreement: "but direct set prediction are postprocessing-free" (p. 3, Sec. 2.1); singular subject with plural verb.
- p. 3 2.1 Set Prediction
direct set prediction are postprocessing-free
Judge's note: Viewed page 3. The next sentence begins 'They need global inference schemes', so 'direct set prediction methods are' was probably intended; either way the printed sentence is ungrammatical.
Second check: none (judge only)
Report a problem with this finding
Typo "the prohibitive inference cost (proportional to output length, and hard to batch) lead to the development of parallel sequence generation" (p. 4, Sec. 2.2); should be "led" (or "leads").
- p. 4 Section 2.2, second paragraph
hard to batch) lead
- p. 4 2.2
prohibitive inference cost (proportional to output length, and hard to batch) lead
Judge's note: Viewed page 4. Singular subject 'cost' with 'lead'; context is past tense ('were first used'), so 'led' is the evident fix.
Second check: none (judge only)
Report a problem with this finding
Typo Agreement: "Recent work [52] demonstrate that the final performance of these systems heavily depends on..." (p. 4, Sec. 2.3); should be "demonstrates".
- p. 4 2.3
Recent work [52] demonstrate that
Judge's note: Viewed page 4; printed as quoted. Single citation [52], singular 'work'.
Second check: none (judge only)
Report a problem with this finding
Typo Reference 4 title is printed "Soft-NMS improving object detection with one line of code" with a wide blank gap where a separator (dash) has been dropped.
- p. 17 Reference 4
Soft-NMS
- p. 17 Reference 4
Soft-NMS improving object
- p. 4 Sec 2.3, Set-based loss
Learnable NMS methods [16,4] and relation networks [17]
Judge's note: The crop and the page-17 image show a gap about two spaces wide after 'Soft-NMS' and no dash or colon, so a character was lost in typesetting; this is on the page, not an extraction artefact. That the lost character is a dash rests on the original title, which is outside the files. The separate claim that Soft-NMS [4] is wrongly cited among 'Learnable NMS methods [16,4]' (page 4) rests on outside knowledge of what that paper does and is not confirmed here.
Second check: none (judge only)
Report a problem with this finding
Typo Equation 2 typesets the predicted box as b̂ with subscript σ̂ followed by "(i))" on the baseline: "L_box(b_i, b̂_σ̂(i))" with the index (i) outside the subscript; it should be b̂_{σ̂(i)}) as in the first term p̂_{σ̂(i)}(c_i) and in the L_match definition.
- p. 5 Equation 2
LHungarian(y, ˆy) =
Judge's note: Viewed the crop and page 5. In Eq. 2 the first term has σ̂(i) fully in the subscript of p̂, but in the box-loss term only σ̂ is subscripted and '(i)' sits at full size on the baseline. A typesetting slip (misplaced brace), meaning is still clear.
Second check: none (judge only)
Report a problem with this finding
Typo Stray full stop mid-sentence: "from C to a smaller dimension d. creating a new feature map z0" (p. 6, Transformer encoder); should be a comma.
- p. 6 Section 3.2, Transformer encoder
to a smaller dimension d. creating a
Judge's note: Viewed page 6: a period follows 'd' and the sentence continues in lower case with 'creating'. Present in the image, not an extraction artefact.
Second check: none (judge only)
Report a problem with this finding
Typo Missing word: "We found helpful to use auxiliary losses [1] in decoder during training" (p. 7); should be "We found it helpful".
- p. 7 Auxiliary decoding losses
We found helpful to use auxiliary losses
Judge's note: Viewed the crop from page 7; printed as quoted.
Second check: none (judge only)
Report a problem with this finding
Typo The detector name is written three ways: "Faster R-CNN" in the text and the Table 1 caption, "Faster RCNN-FPN" (etc.) in the Table 1 rows, and "Faster-RCNN+" in A.4 Baseline.
- p. 9 Table 1 rows
Faster RCNN-FPN
- p. 23 A.4 Baseline
Faster-RCNN+
Judge's note: Read page 9 (caption says "Faster R-CNN", rows say "Faster RCNN-DC5", "Faster RCNN-FPN"...) and page 23 ("Our enhanced Faster-RCNN+ baselines"). Page 10 also has "Faster R-CNN-FPN". Real but purely cosmetic naming inconsistency; the table form may be deliberate shortening.
Second check: none (judge only)
Report a problem with this finding
Typo The operation count is written inconsistently: "FLOPs" (p. 8, "A full comparison of FLOPs"), "FLOPS" (Table 1 caption "to measure FLOPS and FPS", header "GFLOPS/FPS", A.3 "FLOPS computation") and "FLOP count" (p. 10).
- p. 9 Table 1 caption
to measure FLOPS and FPS
- p. 9 Table 1 header
GFLOPS/FPS
Judge's note: Checked pages 8, 9, 10 and 23. The GFLOPS header versus FLOPS in the caption is not a defect (G is only the giga prefix). What remains is inconsistent capitalisation FLOPs/FLOPS for a count of operations, which is cosmetic. "flop_count_operators" on page 23 is the name of a Detectron2 tool in monospace, not an error.
Second check: none (judge only)
Report a problem with this finding
Typo Method name is "UPSNet" / "UPSNet-M" in the Table 5 caption but "UPSnet" / "UPSnet-M" in the table rows; also "In table 5" in lower case and "Faster RCNN-DC5" in Table 1 against "Faster R-CNN" in the text.
- p. 16 Table 5 caption
UPSNet uses the
- p. 16 Table 5, rows 2-3
UPSnet-M
- p. 16 Table 5 row label
UPSnet
- p. 16 Table 5 caption
UPSNet [51]
Judge's note: Page-16 image: caption prints UPSNet three times, rows print UPSnet and UPSnet-M. 'In table 5' is lower case while 'Figure 9' in the same sentence is capitalised. Table 1 row labels 'Faster RCNN-...' confirmed from the text layer against 'Faster R-CNN' throughout the prose. All cosmetic.
Second check: none (judge only)
Report a problem with this finding
Typo Misspelling: "FFN inside tranformers can be seen as 1 × 1 convolutional layers" should read "transformers".
- p. 11 Section 4.2, Importance of FFN
FFN inside tranformers
- p. 11
FFN inside tranformers can be seen as 1 × 1 convo-
Judge's note: Read the paragraph "Importance of FFN" on page 11; the page image shows "tranformers".
Second check: none (judge only)
Report a problem with this finding
Typo Lowercase "results can be found in table 3" and "Results are presented in table 4" (also "In table 5"), against "Table 1" and "Table 2" elsewhere in the running text.
- p. 12 Importance of positional encodings
results can be found in table 3
- p. 12 Loss ablations
Results are presented in table 4
Judge's note: Read page 12 (both lowercase instances) and pages 9-10 ("Results are presented in Table 1", "(Table 2)"). "In table 5" checked in the text layer only. Cosmetic capitalisation inconsistency.
Second check: none (judge only)
Report a problem with this finding
Typo Run-in heading "Decoder output slot analysis In Fig. 7 we visualize" has no full stop after the heading.
- p. 14 4.3
Decoder output slot analysis In Fig. 7
Judge's note: Page-14 image: the bold heading runs straight into 'In Fig. 7'. The next run-in heading on the same page ('Generalization to unseen numbers of instances.') and those on page 16 ('Training details.', 'Main results.') end in a full stop.
Second check: none (judge only)
Report a problem with this finding
Typo Comma splice: "Experimentally, these two approaches give similar results, we report results using the latter method".
- p. 15 4.4
give similar results, we report results
Judge's note: Read on the page-15 image. Two independent clauses joined by a comma. Punctuation only.
Second check: none (judge only)
Report a problem with this finding
Typo Misspelling: "our unified panoptic segmenation approach" (should be 'segmentation').
- p. 16 Section 4.4, Main results
panoptic segmenation approach
Judge's note: Visible on the page-16 image in the Main results paragraph.
Second check: none (judge only)
Report a problem with this finding
Typo Table 5 caption lacks a full stop: "on the COCO val dataset We retrained PanopticFPN".
- p. 16 Table 5 caption
on the COCO val dataset We retrained PanopticFPN
- p. 16 Table 5 caption
on the COCO val dataset We retrained
Judge's note: Page-16 image shows 'dataset We retrained' with no full stop between the sentences.
Second check: none (judge only)
Report a problem with this finding
Typo "we first filter out the detection with a confidence below 85%": singular 'detection' where the plural is meant.
- p. 16 Training details
filter out the detection with a confidence below 85%
Judge's note: Read on the page-16 image, Training details paragraph. Grammar slip.
Second check: none (judge only)
Report a problem with this finding
Typo The method is written "Panoptic FPN [18]" in the first sentence of the Table 5 caption and "PanopticFPN" in the next sentence, the table rows and the text.
- p. 16 Table 5 caption
Panoptic FPN [18]
- p. 16 Table 5 rows
PanopticFPN++
Judge's note: Page-16 image confirms both spellings within the same caption. Overlaps with the naming points in C037 but concerns a different name, so kept as a separate cosmetic observation.
Second check: none (judge only)
Report a problem with this finding
Typo Acknowledgements: "for discussions and advices" ('advice' is uncountable).
- p. 17 Acknowledgements
discussions and advices
Judge's note: Read on the page-17 image.
Second check: none (judge only)
Report a problem with this finding
Typo Reference 19 prints "Rother, C., Dollar, P." without the accent, while references 13, 14, 18, 22, 23 and 24 print "Dollár, P.".
- p. 18 Reference 19
Rother, C., Dollar, P.: Panoptic segmentation
- p. 18 Reference 19
Rother, C., Dollar, P.
Judge's note: Page-18 image: entry 19 has 'Dollar', all other entries with this author have 'Dollár'. Internal inconsistency visible on the page.
Second check: none (judge only)
Report a problem with this finding
Typo Reference 40 prints the author as "Leal-Taix, l.:" where reference 39 prints "Leal-Taixé, L.".
- p. 19 Reference 40
Leal-Taix, l.:
- p. 19 Reference 39
Leal-Taixé,
- p. 19 Reference 40
Leal-Taix, l.
- p. 19 Reference 40
Leal-Taix, l.: Deepsetnet
Judge's note: The crop of page 19 shows 'Leal-Taix, l.:' in entry 40 (missing final letter, lower-case initial) directly below 'Leal-Taixé, L.' in entry 39. It is in the rendered page, not an extraction artefact.
Second check: none (judge only)
Report a problem with this finding
Typo "shown in Fig.11" and "shown in Fig.12" lack the space used elsewhere ("Fig. 7", "Fig. 10").
- p. 24 A.5
shown in Fig.12
- p. 23 A.5
shown in Fig.11
Judge's note: Viewed pages 23 and 24: both references print without a space. A text search shows the other in-text references use 'Fig. 4', 'Fig. 6', 'Fig. 7', 'Fig. 10' with a space. Spacing inconsistency only.
Second check: none (judge only)
Report a problem with this finding
Typo "the number of object per class" should read "objects".
- p. 24 A.5
the number of object per class
Judge's note: Viewed page 24; the text is printed as quoted.
Second check: none (judge only)
Report a problem with this finding
Typo "the experiments suggests that the model" has a subject-verb agreement error ("suggest").
- p. 24 A.5
the experiments suggests that the model
Judge's note: Viewed page 24; the text is printed as quoted.
Second check: none (judge only)
Report a problem with this finding
Typo The Fig. 12 caption ends "misses more and more objects" with no full stop.
- p. 25 Fig. 12 caption
misses more and more objects
Judge's note: Viewed page 25: the caption's last sentence has no final period, although its earlier sentences do. The Fig. 10 caption (page 22) ends with a period; the Fig. 11 caption (page 24) also ends without one, so the paper is not consistent. I did not check the Fig. 5 caption that the reviewer also mentions.
Second check: none (judge only)