DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection
1. Paper overview
In one sentence: DINO improves object detection by choosing image-specific anchor positions, teaching noisy anchors when to reject objects, and sharing box supervision across adjacent decoder layers. e01e04e05e07e08e10e11e13
| At a glance | What to know |
|---|---|
| Research problem | Source description DETR removes hand-designed detection stages through set prediction and bipartite matching, but converges slowly. DINO addresses unstable matching and ambiguous anchor selection while testing whether this detector family can compete with optimized systems at larger model and data scales. e02e04e12 |
| Core mechanism | Source description Contrastive denoising teaches nearby perturbed anchors to reconstruct objects and farther perturbed anchors to predict background. e05 |
| A key reported result | COCO val2017 object detection with longer training: 24 epochs: 50.4 (four scales), 51.3 (five scales). 36 epochs: 50.9 and 51.2. Box AP. ResNet-50; COCO train2017; DINO trained for 24 or 36 epochs. DN-Deformable-DETR reaches 48.6 after 50 epochs; the 24-epoch gains are 1.8 and 2.7 AP points. This supports faster convergence by epoch count, not a matched-time experiment. Five-scale AP does not improve from 24 to 36 epochs. e09e11 |
| Reading caution |
Core contributions
- Source description
Contrastive denoising teaches nearby perturbed anchors to reconstruct objects and farther perturbed anchors to predict background. e05
- Author claim
DINO establishes a leading COCO detector in the authors’ comparison, including a Swin-L model pretrained on Objects365. This leaderboard claim is historical and specific to the compared systems. e12
Figure 2. Encoder evidence supplies anchor positions and attention memory; a separate denoising branch supplies training supervision. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the feature maps on the left and follow the flattening arrows into the encoder. The encoder outputs have two destinations: the decoder’s key/value input and the query-selection block. Follow the latter’s long arrow to the decoder’s initial anchors. The content queries enter separately as learnable parameters; the caption explicitly distinguishes them from selected encoder features. On the far right, positive and negative ground-truth perturbations occupy the CDN branch, beside the matching queries. The vertical repetition labels N and M denote encoder and decoder layer counts. The drawing combines the ordinary prediction pathway with the additional training pathway. e04e05e07
What it supports. DINO separates where a query starts from the content it will gather. Image-conditioned anchors provide spatial guidance, while decoder attention builds content for classification. The noisy-ground-truth branch supervises that shared detector during training; the source describes final inference outputs as object boxes and classes.
Where the evidence stops. Figure 2 is an architectural overview, not a complete attention-mask or loss specification. Ground-truth inputs belong to denoising training. The diagram contains no temporal rollout, action prediction or interaction with an environment.
2. Motivation
2.1 The problem and the proposed response
DETR removes hand-designed detection stages through set prediction and bipartite matching, but converges slowly. DINO addresses unstable matching and ambiguous anchor selection while testing whether this detector family can compete with optimized systems at larger model and data scales. e02e04e12
2.2 What this reading follows
DINO asks how a Transformer detector can learn accurate object sets quickly while remaining competitive at larger scale. Its answer combines three changes to the query pipeline: initialize positions from the image, train on both reconstructable and rejectable noisy anchors, and let later box predictions supervise an adjacent earlier refinement. The figures below connect those operations to a diagnostic and two quantitative comparisons. Read the results with their configurations attached: the longer-training comparison uses unequal epoch budgets, and the component ablation has a different endpoint from the revised main table. This edition reviews the supplied July 2022 arXiv v4, not a separately verified ICLR 2023 edition. e01e04e05e07e08e10e11e13
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | Foundational work |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
This table preserves the labels recorded at reading time. The current major category is Related resources. View the current classification.
3.1 Evidence-based assessment
Supports the recorded classification
The recorded foundational visual-representation placement fits an object detector. Its encoder–decoder predicts boxes and classes, not future world states or actions. The world-action-model architecture, prediction-paradigm and quadrant fields are therefore not applicable; a single detector does not establish a One Model world-action architecture. e04e09
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Let the image choose the anchor’s position
Begin with the separation between position and content in a decoder query. An anchor box gives the decoder a spatial starting point, but the content representation still has to gather useful image evidence. Pure query selection derives both components from selected encoder features. DINO instead uses the selected positions to initialize anchors while leaving content queries as learned parameters. Figure 5 makes “static” precise: the initial content is the same across images at inference, although decoder attention subsequently makes its representation image-dependent. The authors argue that selected encoder content may cover part of an object or multiple objects. Reader interpretation: this division trusts encoder localization as a starting cue without forcing the decoder to inherit all of its preliminary semantic ambiguity. e03e04e07
Figure 3. Each denoising group pairs reconstruction targets with nearby background targets. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each group from the query tokens at the bottom through the shared decoder to the targets at the top. Yellow positives lead to reconstruction loss; brown negatives lead to “no object.” The concentric squares at right are a conceptual view of perturbation magnitude around a ground-truth box. Section 3.3 specifies positive noise below λ1 and negative noise greater than λ1 but below λ2, consistent with the inner and outer regions in the caption. The actual anchors have four coordinates; this two-dimensional illustration does not describe literal square regions on the input image. Multiple groups repeat this supervision with perturbed examples. e05e21e22
What it supports. The contrast is between two supervised responses to noisy anchors: recover an object or reject the candidate. Positive box reconstruction uses L1 and GIoU plus focal classification loss; negative queries use background focal loss. This extends reconstruction-only denoising with an explicit rejection target.
Where the evidence stops. The graphic does not show a feature-similarity contrastive objective. Its spatial analogy is schematic. Appendix D.3’s noise radii and Table 8’s noise-scale entry are not explicitly reconciled, so the drawing alone cannot determine a reproduction’s sampling configuration.
5.2 Teach the anchor when reconstruction is the wrong response
Denoising starts from supervision unavailable at inference: a ground-truth box and label. Perturbing that box creates a nearby anchor whose intended target is known, easing training. DINO adds a second, more strongly perturbed anchor with a background target. Figure 3 should therefore be read as a supervised distinction between recovering an object and rejecting a candidate, rather than as a generic embedding-similarity loss. Figure 4 then asks whether the initial anchors associated with successful matches lie closer to their objects, especially small ones. Reader interpretation: the diagnostic is compatible with the proposed rejection mechanism but is not sufficient to prove it. Table 7 helps constrain the alternative that improvement merely comes from more queries: 100 CDN pairs outperform 200 ordinary DN queries at equal total count. e05e06e18
Figure 6. Look forward twice changes the supervised gradient route while retaining detached refinement inputs. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the dashed-arrow legend first: a dashed connection preserves the forward box value but detaches its gradient. In both panels, follow the lower sequence from an undetached refined box b_i′ to its detached counterpart b_i and then to the next layer’s offset. The right panel separates the supervised predictions from that refinement sequence. Equation (2) resolves the exact operation: the prediction at layer i combines the preceding undetached box b_(i−1)′ with offset Δb_i. That distinction lets the adjacent prediction loss affect an earlier refinement. The equation and footnote, rather than the schematic arrow geometry, define the normalized-coordinate box update. e08e13
What it supports. “Twice” describes reuse of a predicted offset in the training graph. The next supervised prediction can improve the preceding box estimate as well as its own offset. It does not prescribe running the complete detector twice or eliminating all gradient detachment between layers.
Where the evidence stops. The claimed benefit is conditional on the training design. Table 4 measures this addition after mixed query selection, rather than across every component combination. The source’s Update operation uses inverse-sigmoid/sigmoid coordinates, not raw coordinate addition.
5.3 Separate the refinement value from the path that trains it
Look forward twice is easiest to understand by distinguishing numerical boxes from their gradient history. Detaching a refined box leaves its value unchanged, allowing the next decoder layer to continue refinement without backpropagating through that input. DINO retains that route, but constructs the supervised prediction with the previous undetached box and the current layer’s offset. Equation (2) expresses the distinction directly. A later prediction loss can consequently improve the preceding refinement as well as the current offset. Reader interpretation: this is a local extension of credit assignment across adjacent layers, not temporal prediction. Table 4 reports an improvement after adding this route to mixed query selection, but it does not establish the same gain for every baseline or show that removing all detachment would work. e08e13
5.4 Training and inference
During training
The base setup uses six encoder and six decoder layers, hidden dimension 256, 900 matching queries, AdamW, batch size 16, learning rate 10⁻⁴ and weight decay 10⁻⁴. Learning rate drops tenfold at epochs 11, 20 or 30 for the 12-, 24- or 36-epoch schedules. ResNet-50 starts from ImageNet-1k weights. e09e21e22
L1 and GIoU box losses have coefficients 5 and 2; focal classification loss has coefficient 1, α=0.25 and γ=2. Auxiliary decoder and query-selection losses supplement matching supervision. Dynamic denoising groups control query count; the default description uses 100 positive/negative pairs. Decoder prediction heads share parameters. e05e19e21
Four scales use backbone stages 2–4 plus a downsampled stage-4 map; five scales add stage 1. Training uses random crops and resizing, with shorter side 480–800 and longer side at most 1333 pixels. e21
During inference
An image passes through encoding, query initialization and iterative decoding to produce boxes and classes. Ground-truth denoising supplies training supervision. Ordinary detection is free of RPN and NMS and executes no environmental action. e04e05e12
Optional Swin-L test-time augmentation uses scales and horizontal flips. A main augmentation selects maximum-IoU partners above a threshold in other views, followed by score/weight aggregation. These results are distinct from ordinary inference. e15
5.5 Implementation flow
- Encode spatial evidence
A ResNet-50 or Swin-L backbone extracts multi-scale features. Flattened features and positional embeddings enter the Transformer encoder. A positional query is a box (x, y, w, h): center coordinates, width and height. e03e04
- Initialize positions, then gather content
Top-K encoder selections initialize anchor boxes. Content queries remain learned parameters shared across images at initialization. Deformable attention gathers encoder evidence and refines queries through the decoder; classification uses refined content features. e04e07
- Learn reconstruction and rejection
Each ground-truth box generates positive and negative denoising examples during training. Positive noise is below λ1; negative noise lies between λ1 and λ2. Positives receive box reconstruction and classification supervision; negatives receive background classification supervision. e05
- Propagate adjacent supervision
Look forward twice preserves a detached box for the next refinement layer but forms the supervised prediction with the preceding undetached box. The next layer’s prediction loss can therefore train the preceding box refinement. e08
6. Experiments & results
DINO turns image features into object boxes and classes through dynamic anchor queries. Contrastive denoising, mixed query initialization and adjacent-layer box supervision improve DETR-style detection, with an accuracy–computation tradeoff between four and five feature scales. This is the July 2022 v4 detector paper; it studies neither future-state prediction nor action execution.
6.1 Read the original evidence
Table 2. DINO exceeds the listed DN baseline with fewer training epochs, with separate four- and five-scale configurations. Original paper, p. 13 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the Epochs column before the bold AP values. The DN-Deformable-DETR reference row is a 50-epoch model, whereas DINO’s first two rows use 24 epochs and its final two use 36. Within each DINO pair, distinguish four scales from five scales. The parenthetical AP increments use the 48.6 baseline; they are AP-point differences, not relative percentages. AP50 and AP75 report the table’s IoU-specific measures, while the S, M and L subscripts separate object sizes. The caption fixes the shared backbone and evaluation split as ResNet-50 and COCO val2017, but does not make the training budgets identical. e09e11e10
What it supports. At 24 epochs, four-scale DINO reaches 50.4 AP and five-scale DINO 51.3, compared with 48.6 for the 50-epoch DN baseline. The five-scale score is 51.2 at 36 epochs, so longer training does not produce a monotonic gain for every configuration in this table.
Where the evidence stops. Epoch count is not elapsed compute. Table 1 reports a substantial throughput difference between DINO’s four- and five-scale models. Table 2 provides no variability estimates, so its small within-family differences should not be interpreted as statistically established effects.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| COCO val2017 object detection after 12 epochs COCO train2017; ImageNet-1k-pretrained ResNet-50; 900 DINO queries; four or five feature scales. | Four scales: 49.0 / 32.0, 279 GFLOPS, 24 FPS. Five scales: 49.4 / 32.3, 860 GFLOPS, 10 FPS. Box AP / small-object AP; GFLOPS and FPS | Four-scale DN-Deformable-DETR: 43.4 / 24.8, 265 GFLOPS, 23 FPS. AP gains are 5.6 and 6.0 points. Four-scale DINO improves accuracy at similar reported throughput. The fifth scale gives 0.4 more AP at greater cost. Timing uses the same A100 GPU; other rows vary in queries and feature configurations. e09e10 |
| COCO val2017 object detection with longer training ResNet-50; COCO train2017; DINO trained for 24 or 36 epochs. | 24 epochs: 50.4 (four scales), 51.3 (five scales). 36 epochs: 50.9 and 51.2. Box AP | DN-Deformable-DETR reaches 48.6 after 50 epochs; the 24-epoch gains are 1.8 and 2.7 AP points. This supports faster convergence by epoch count, not a matched-time experiment. Five-scale AP does not improve from 24 to 36 epochs. e09e11 |
| COCO detection after large-scale pretraining Swin-L pretrained on ImageNet-22k; detector pretrained on Objects365 then fine-tuned on COCO at larger image scale. | Without TTA: 63.1 / 63.2. With TTA: 63.2 / 63.3; 218M parameters. Box AP on val2017 / test-dev | SwinV2-G reports 62.5 / 63.1 with TTA and 3.0B parameters. The headline 63.3 uses TTA. Different backbones, pretraining and annotation use prevent an isolated architectural comparison. e12e15e20 |
| Incremental component ablation Table 4’s optimized DN-DETR baseline; the table does not independently specify its complete training configuration. | Optimized DN-DETR 44.9 → pure query selection 46.5 → mixed selection 47.0 → look forward twice 47.4 → CDN 47.9. Box AP; small-object AP | Original DN-DETR: 43.4. Adding CDN raises small-object AP from 29.9 to 31.2. Sequential gains support complementary components but do not measure all interactions. The 47.9 endpoint differs from the revised main table’s 49.0/49.4. e13e10 |
| Denoising query-budget diagnostic ResNet-50 on COCO validation; dynamic denoising groups; Table 7. | 100 CDN pairs (200 queries): 47.9; 200 DN queries: 47.4; 1000 DN queries: 47.6; no DN: 45.1. Box AP | CDN exceeds DN at equal total query count by 0.5 AP point. More reconstruction queries alone do not explain CDN’s reported gain. No run-to-run uncertainty is supplied. e18e19 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 4. A tail-distance diagnostic makes the small-object anchor-selection hypothesis more specific. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the legend independently in each panel: yellow is DN and blue is DINO. Panels (a) and (b) report ATD(100), so lower is better: Equation (1) averages the 100 largest L1 distances between ground-truth boxes and the initial anchors whose final predictions match them. Panel (a) pools all objects; panel (b) restricts to small objects. Panel (c) changes metric and epoch range, showing small-object AP, where higher is better. Its vertical axis uses fractional values, unlike the AP-point notation in the tables. Compare trends within panels rather than treating their different vertical scales as directly comparable. e06e05
What it supports. The all-object distance curves are similar, while the small-object diagnostic more often favors DINO. The accompanying text reports a 1.3 AP-point small-object gain over DN in the 12-epoch ResNet-50 setting. The evidence focuses the authors’ explanation on anchor quality for small objects rather than uniform improvement across all anchors.
Where the evidence stops. ATD is computed after matching final predictions and measures a selected tail of initial-anchor distances. It is not final-box error or a direct duplicate count. These curves show no uncertainty bands and do not prove that ATD causally mediates the AP gain.
Table 4. The component table separates engineering gains from the three proposed changes. Original paper, p. 15 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow the row descriptions and component columns together. QS means query selection, CDN means contrastive denoising and LFT means look forward twice. Row 2 first improves DN-DETR with pipeline engineering. Row 3 adds pure query selection and forms the strong baseline. Row 4 changes that selection to mixed, row 5 adds LFT, and row 6 adds CDN; the checkmarks agree with those descriptions. Read the AP column for overall performance, then inspect AP_S: mixed selection raises it, LFT lowers it in the next row, and CDN raises it again. This is a sequential construction, not a complete grid of independent interventions. e13e10
What it supports. Beyond the 46.5-AP strong baseline, mixed selection adds 0.5 point, LFT adds 0.4 and CDN adds 0.5, reaching 47.9. CDN also raises AP_S from 29.9 to 31.2 in the final step. The table therefore supports incremental benefit in this particular component order.
Where the evidence stops. The source does not fully map this table’s 47.9 endpoint to the revised main table’s 49.0/49.4 configurations. Do not transfer its increments to those endpoints as an exact decomposition. Missing component combinations and uncertainty estimates limit causal interpretation.
7. Analysis & limitations
7.1 What the evidence leaves open
Results provide no seed-level variability or confidence intervals. Sequential ablation cannot establish a universal independent gain per component. Table 4’s configuration does not fully reconcile its endpoint with revised Table 1. e10e13e18
ATD emphasizes a selected tail of matched anchors; the duplicate-box illustration is one person-detection example. These support a mechanism hypothesis without establishing a general duplicate-error rate or performance outside COCO. e06e14
Appendix D.3 gives λ1=1 and λ2=2, while Table 8 lists dn_box_noise_scale=0.4 without mapping these settings. TTA weights and IoU threshold are unspecified; Equation (3) divides by a count despite a score-weighted numerator. e15e21e22
7.2 Questions for discussion
- Would the small-object ATD improvement survive equal denoising-query and compute budgets? [e06, e18]
- Which interactions would a factorial query-selection/CDN/gradient-routing ablation reveal beyond Table 4’s sequential gains? [e13]
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction needs the stated split, ImageNet initialization, scale choice, shared heads, dynamic denoising and losses. Table 8 adds backbone learning rate 10⁻⁵, clipping 0.1 and FrozenBatchNorm2d; that normalization entry does not establish a frozen backbone. e09e19e21e22
The large setup uses 26 Objects365 epochs on 64 A100 GPUs, then 18 COCO epochs on 16 A100s, one image per GPU. It retains the first 5,000 Objects365 validation images and moves 75,000 into training. Fine-tuning uses up to 1200×2000 images, checkpointing, mixed precision and 1000 DN queries. e20
Table 5 reports approximately 55 minutes/epoch and 16 GB GPU memory for the 47.9-AP, 12-epoch base model on eight A100s, not the revised five-scale or Swin-L models. Appendix C tests depth with DN queries without negatives, rather than full CDN. e16e17
Proposed checks compare CDN and DN at equal query count and isolate the adjacent-layer gradient route. Hold other settings fixed, measure AP and mechanism-specific diagnostics, repeat seeds, and resolve noise/TTA ambiguities before claiming exact reproduction. e08e13e18e21e22e15
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Hold the denoising query budget constant
Reader-proposed check, not an executed experiment: compare 100 CDN pairs with 200 ordinary DN queries using the same ResNet-50, COCO split, 12-epoch schedule, matching-query count, feature scales, mixed selection, LFT and loss coefficients. Record actual padded query counts and memory so nominal equality does not hide different workloads. Repeat at least three seeds and report AP, AP_S and ATD(100) for all and small objects. Also preregister a fixed matching rule for counting duplicate detections. If CDN fails to improve AP_S or duplicate counts at equal budget, the rejection explanation weakens; lower ATD alone would not establish a useful detector improvement. Document how the conflicting noise settings were resolved before comparing outcomes. e05e06e09e18e19e21e22
Check 2: Isolate the adjacent prediction gradient
Reader-proposed check, not an executed experiment: start two models from identical parameters and inputs, retain the detached refinement chain in both, and vary only whether the supervised prediction receives the preceding undetached box or its detached counterpart. Before training, confirm equal forward box values and backpropagate one selected decoder prediction loss to check whether the preceding box-update head receives the extra gradient only in the LFT condition. Then train under a fixed 12-epoch ResNet-50 configuration across repeated seeds, holding query selection and denoising constant. Compare AP and AP75 alongside gradient diagnostics. A correct gradient difference without reproducible AP benefit would verify the implementation while failing to replicate the reported accuracy benefit. e08e09e13e21e22
8.3 Reading coverage
Visual audit: Visually inspected the title/author/version page; Figures 1–8; Tables 1–8; method text, Equations (1)–(3), evaluation protocols and Appendix A–D configuration pages listed here. All six final crops were separately viewed after rendering at 300 DPI; arrows, gradient-detachment marks, noise-region definitions, legends and table columns were checked against captions and equations. Pages 3–4 and 16–18 were read in the complete text but contain no additional retained claim requiring visual verification in this bundle. Separate supplements, other revisions and code remain outside this pass. The artwork was preserved; no chart was reconstructed.
PDF pages inspected for this edition: 1, 2, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 19, 20, 21, 22, 23. Appendix coverage: reviewed.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Abstract
- 1 Introduction
- 2 Related Work (2.1–2.3)
- 3 DINO: DETR with Improved DeNoising Anchor Boxes (3.1–3.5)
- 4 Experiments (4.1–4.4)
- 5 Conclusion
- References
- A Test Time Augmentations (TTA)
- B Training Efficiency
- C Additional Analysis on our Model Components
- D More Implementation Details (D.1–D.4)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- The extraction limitation above was addressed by inspecting original PDF pages and all final crops. All five supplied text chunks, covering all 23 pages, were read; no separately supplied supplement was available.
- Identity and edition scope: the title and all eight authors match the catalog. The title page identifies arXiv:2203.03605v4, 11 July 2022. The catalog records ICLR 2023; the supplied artifact does not establish the venue edition or its equivalence to this revision. Page 12 attributes improved ResNet-50 results relative to the first version to engineering changes.
- Code, linked resources, other revisions and cited papers were not inspected. No experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title, author block and arXiv margin stamp
Exact catalog title; authors Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni and Heung-Yeung Shum. Stamp: arXiv:2203.03605v4, 11 Jul 2022. Four affiliations appear on the title page.
Go to primary source ↓e02PDF p. 2, Section 1, paragraphs on DETR and DN-DETR
Detection is set prediction with bipartite matching; slow convergence motivates dynamic anchors and denoising.
Go to primary source ↓e03PDF p. 5, Section 3.1, first paragraph
Positional queries are four-dimensional boxes (x,y,w,h); content and position are separate query components.
Go to primary source ↓e04PDF p. 6, Figure 2 and caption; p. 7, Section 3.2
Multi-scale features feed the encoder; selected anchors and learnable content queries feed iterative deformable decoding. Final outputs are boxes and classes; denoising is an extra training branch.
Go to primary source ↓e05PDF pp. 7–8, Section 3.3 and Figure 3 with caption
Positive noise is below λ1; negative noise lies between λ1 and λ2. Each ground truth supplies a pair per group. Positive reconstruction uses L1/GIoU and focal losses; negatives use background focal loss. The drawing projects four-dimensional anchors into two dimensions.
Go to primary source ↓e06PDF p. 8, Section 3.3, Effectiveness; p. 9, Equation (1), Figure 4 and accompanying text
ATD averages the largest initial-anchor/ground-truth L1 distances after matching final predictions. Figure 4 compares DN and DINO for all objects, small objects and small-object AP; text reports a 1.3-point small-object AP gain in the 12-epoch ResNet-50 setting.
Go to primary source ↓e07PDF pp. 9–10, Section 3.4, Figure 5 and caption
Mixed selection initializes anchor positions from selected encoder features while content initialization stays learnable and image-independent. Authors argue preliminary selected content can be ambiguous.
Go to primary source ↓e08PDF p. 10, Figure 6; p. 11, Section 3.5, Equation (2) and footnote 3
Detached boxes remain on the refinement path; supervised predictions use preceding undetached boxes. Adjacent losses influence box refinement. Update uses inverse-sigmoid/sigmoid coordinates.
Go to primary source ↓e09PDF p. 11, Section 4.1, Dataset and Backbone
ResNet-50 uses ImageNet-1k initialization and COCO train2017; Swin-L uses ImageNet-22k then Objects365/COCO. Evaluation uses COCO val2017 and, for Swin-L, test-dev.
Go to primary source ↓e10PDF p. 12, Table 1, DINO and DN-Deformable-DETR rows, caption and Section 4.2
Twelve-epoch AP: 49.0/49.4 versus 43.4; small AP 32.0/32.3 versus 24.8. DINO cost: 279/860 GFLOPS and 24/10 FPS, versus 265 and 23 for DN. Caption states query differences; timing uses A100. Text notes engineering improvements over the first paper version.
Go to primary source ↓e11PDF p. 13, Table 2, DN-Deformable-DETR and four DINO rows; caption
DINO four/five scales: 50.4/51.3 AP at 24 epochs and 50.9/51.2 at 36; DN baseline 48.6 at 50 epochs. All use ResNet-50 and COCO val2017.
Go to primary source ↓e12PDF p. 14, Table 3, DINO-SwinL and SwinV2-G rows, caption and Section 4.3
DINO has 218M parameters and 63.1/63.2 val/test AP without TTA, 63.2/63.3 with TTA. SwinV2-G has 3.0B parameters and 62.5/63.1 with TTA. Pretraining and mask use differ. End-to-end is defined as free of RPN/NMS.
Go to primary source ↓e13PDF p. 14, Section 4.4; p. 15, Table 4, rows 1–6 and caption
Sequential AP is 43.4, 44.9, 46.5, 47.0, 47.4, 47.9. Rows 5–6 small AP is 29.9→31.2. QS/CDN/LFT columns identify component additions; optimized baselines include engineering changes.
Go to primary source ↓e14PDF p. 15, Figure 8 and caption
One example compares person detections using DN and CDN; an arrow identifies three duplicate boxes in the DN output.
Go to primary source ↓e15PDF p. 19, Appendix A, Equation (3) and surrounding definitions
Swin-L TTA uses scales and horizontal flips, main-view IoU association and a score/weight aggregation formula. The threshold and augmentation weights are named but not numerically supplied; the printed denominator is the sum of inclusion indicators.
Go to primary source ↓e16PDF p. 19, Appendix B; p. 20, Table 5, DINO row and caption
ResNet-50 base training uses eight A100 GPUs, two images per GPU, approximately 55 min/epoch, 16 GB memory, 12 epochs and 47.9 AP. Baseline settings are separately qualified in the caption.
Go to primary source ↓e17PDF p. 20, Table 6 and Appendix C depth analysis
Layer-count ablation uses ResNet-50, COCO val2017, 12 epochs and 100 DN queries without negatives. Six/six layers give 47.4 AP; six/two give 44.4.
Go to primary source ↓e18PDF p. 20, Table 7 and caption; p. 21, Appendix C continuation
100 CDN pairs contain 200 queries and give 47.9 AP; 200 DN gives 47.4, 1000 DN gives 47.6, no DN gives 45.1. Dynamic denoising groups are used.
Go to primary source ↓e19PDF p. 21, Appendix D.1
DINO fixes the denoising query budget and adjusts group count according to object count to reduce padding-related memory overhead.
Go to primary source ↓e20PDF p. 21, Appendix D.2; p. 22, Appendix D.3, Training augmentation
Objects365 validation is repartitioned into first 5,000 retained and remaining 75,000 added to training. Pretraining/fine-tuning: 26/18 epochs on 64/16 A100 GPUs, local batch one. Larger fine-tuning resolution, checkpointing, mixed precision and 1000 DN queries are stated.
Go to primary source ↓e21PDF pp. 21–22, Appendix D.3, Basic hyper-parameters, Loss function, Detailed model components, Training augmentation and Multi-scale setting
Six encoder/decoder layers, dimension 256, 900 queries, AdamW and schedule are specified. Loss coefficients are 1/5/2, focal α=0.25 and γ=2. Shared heads, random crop/resize, scale construction, 100 CDN pairs and λ1=1, λ2=2 are described.
Go to primary source ↓e22PDF p. 23, Table 8, optimizer, normalization, cost and noise rows
Backbone learning rate 1e-5, clipping 0.1 and FrozenBatchNorm2d are listed. Matching class cost is 2 while classification loss coefficient is 1. dn_box_noise_scale=0.4 and dn_label_noise_ratio=0.5 appear without a mapping to the prose noise radii.
Go to primary source ↓8.5 Primary sources
DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection ↗
PDF · 9,689 extracted words
Source fingerprint
2cda48e74a49e3552aedbf61bc6436767ed3867cb710bbd2b79e48624fd66126