Source-linked AI summary

DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection

Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, Heung-Yeung Shum

arXiv:2203.03605v4cs.CV

TL;DR

DETR-like detectors needed better convergence, performance, and evidence of scalability to large models and datasets. DINO combines improved denoising training, mixed query selection, and look forward twice, achieving state-of-the-art COCO results while using smaller model and pre-training data sizes than compared leaderboard models. Its results establish an end-to-end Transformer detector as a strong detection framework.

  • Problem

    DETR-like models were inferior to improved classical detectors and lacked reported scalability results for large backbones and datasets.

  • Method

    DINO is an end-to-end DETR-like detector using contrastive denoising training, mixed query selection, and look forward twice for box prediction.

  • Results

    DINO achieves the best reported results of 63.2 AP on COCO val2017 and 63.3 AP on test-dev, while improving ResNet-50 results over previous best DETR-like models.

  • Takeaways & Limitations

    DINO establishes an end-to-end Transformer detector as a mainstream detection framework with strong performance and scalability.

Abstract

from arXiv · show

We present DINO (\textbf{D}ETR with \textbf{I}mproved de\textbf{N}oising anch\textbf{O}r boxes), a state-of-the-art end-to-end object detector. % in this paper. DINO improves over previous DETR-like models in performance and efficiency by using a contrastive way for denoising training, a mixed query selection method for anchor initialization, and a look forward twice scheme for box prediction. DINO achieves $49.4$AP in $12$ epochs and $51.3$AP in $24$ epochs on COCO with a ResNet-50 backbone and multi-scale features, yielding a significant improvement of $\textbf{+6.0}$\textbf{AP} and $\textbf{+2.7}$\textbf{AP}, respectively, compared to DN-DETR, the previous best DETR-like model. DINO scales well in both model size and data size. Without bells and whistles, after pre-training on the Objects365 dataset with a SwinL backbone, DINO obtains the best results on both COCO \texttt{val2017} ($\textbf{63.2}$\textbf{AP}) and \texttt{test-dev} (\textbf{$\textbf{63.3}$AP}). Compared to other models on the leaderboard, DINO significantly reduces its model size and pre-training data size while achieving better results. Our code will be available at \url{https://github.com/IDEACVR/DINO}.

1 Introduction

DINO targets the performance and scalability gaps of DETR-like detectors with improved denoising training, query initialization, and box prediction. It achieves strong COCO results at short training schedules and with larger models and datasets.

  • DETR background: DETR removes hand-designed components by formulating object detection as set prediction with bipartite matching and learnable queries.Queries probe object existence and combine image features similarly to soft ROI pooling.
  • Motivation: DETR-like models previously faced slow convergence and unclear query meaning, while DAB-DETR and DN-DETR addressed these issues with dynamic anchor boxes and denoising.Their combination made DETR-like models competitive with classical detectors in training efficiency and inference performance.
  • Motivation: Previous DETR-like models remained below 50 AP on COCO, and their scalability to large backbones and datasets had not been reported.DINO explicitly aims to address both performance and scalability concerns.
  • DINO: DINO improves denoising training, query initialization, and box prediction through contrastive denoising, mixed query selection, and look forward twice.The model builds on DN-DETR, DAB-DETR, and Deformable DETR.
  • Results: 49.4 AP in 12 epochs and 51.3 AP in 24 epochs with ResNet-50 and multi-scale features improved over the previous best DETR-like model by +6.0 AP and +2.7 AP.DINO also achieved 63.2 AP on COCO val2017 and 63.3 AP on test-dev after Objects365 pre-training with SwinL.

2 Related Work

Related work spans classical convolutional detectors, end-to-end Transformer detectors, query-focused DETR variants, and large-scale pre-training. These directions motivate DINO's emphasis on end-to-end detection, query design, and scalability.

  • Convolution-based detectors: Classical detectors include two-stage and one-stage convolutional models built around hand-crafted anchors or reference points.Two-stage models use region proposal networks, while one-stage models predict offsets relative to predefined anchors.
  • Transformer detectors: DETR introduced end-to-end Transformer detection without hand-designed anchor design and non-maximum suppression, prompting work on faster convergence.Follow-up approaches include encoder-only designs and dynamic decoders focused on important regions.
  • Decoder queries: Query-focused methods associate decoder queries with spatial information using reference points, dense encoder predictions, or dynamic anchor boxes.Deformable DETR, Efficient DETR, and DAB-DETR represent successive approaches to spatially informed queries.
  • Large-scale pre-training: Large-scale pre-training has improved computer-vision models, with leading detectors using large backbones and substantial pre-training datasets.The related work contrasts these resource-intensive approaches with DINO's publicly available pre-training setting.

3 DINO: DETR with Improved DeNoising Anchor Boxes

DINO improves DETR-like detection through contrastive denoising, mixed query selection, and look forward twice box prediction. Its architecture combines dynamic anchor-box queries, deformable attention, and image-adaptive positional initialization.

  • Model overview: DINO formulates decoder positional queries as dynamic 4D anchor boxes and refines them across decoder layers.The boxes use center coordinates and dimensions as query representations.
  • Contrastive DeNoising Training: Contrastive DeNoising adds positive and negative anchor queries so the decoder learns to reject anchors without nearby objects.Positive and negative examples are defined by their distance from a ground-truth box, and multiple CDN groups are used during training.
  • Contrastive DeNoising Training: +1.3 AP on small objects is achieved by CDN queries over DN queries after 12 epochs with ResNet-50 and multi-scale features.CDN finds better anchors for small objects, while DN is described as sufficient for selecting good anchors overall.
  • Mixed Query Selection: Mixed query selection initializes anchor boxes from positional information in the selected top-K encoder features while keeping content queries static or learnable.This differs from Deformable DETR, which uses selected features to enhance both positional and content queries.
  • Look Forward Twice: Look forward twice updates box predictions so later-layer box information can influence adjacent earlier-layer parameters.The method contrasts with Deformable DETR’s look forward once update, which blocks gradient backpropagation for stabilization.

4 Experiments

DINO is evaluated on COCO across training schedules, backbones, model scales, and ablations. It improves detection performance and convergence while scaling to large models and datasets.

  • Experimental Setup: DINO is evaluated on COCO 2017 using ResNet-50 or SwinL backbones, with SwinL models pre-trained on Objects365 before COCO fine-tuning.The ResNet-50 setting uses train2017 without extra data, while SwinL uses ImageNet-22k pre-training followed by Objects365 pre-training.
  • Main Results: DINO uses 900 queries, while most other DETR-like models use 300 queries in the 12-epoch comparison.The table compares models on COCO val2017 and reports GFLOPS and FPS for models without multi-scale features using ResNet-50-DC5.
  • Main Results: +5.6 AP is obtained with ResNet-50 and 4-scale features, while +6.0 AP is obtained with 5-scale features under the same setting.The 4-scale model adds little computation and parameter overhead, and gains +7.2 AP or +7.5 AP on small objects with 4 or 5 scales, respectively.
  • Main Results: +1.8 AP and +2.7 AP are achieved after 24 epochs with 4 and 5 scales, respectively; after 36 epochs, gains reach +2.3 AP and +2.6 AP.The convergence comparison is shown in Fig. 7, supporting evaluation of both training speed and final performance.
  • Comparison with SOTA Models: 63.2 AP on COCO val2017 and 63.3 AP on test-dev establish DINO as an end-to-end Transformer detector with strong scalability to larger model and data sizes.The SwinL model uses Objects365 pre-training and achieves these results without test-time augmentation.
  • Ablation Studies: Ablation studies evaluate the effectiveness of DINO’s proposed algorithm components against optimized DN-DETR and query-selection baselines.The components are reported as Query Selection, Contrastive De-Noising Training, and Look Forward Twice.

5 Conclusion

DINO combines several training and prediction improvements into an end-to-end Transformer detector. The paper reports improved efficiency and detection performance, while its test-time augmentation exploration remains limited.

  • Conclusion: DINO combines contrastive denoising training, mixed query selection, and look forward twice to improve training efficiency and final detection performance.The model is presented as an end-to-end Transformer detector.
  • Conclusion: DINO achieves 63.3 AP on COCO 2017 test-dev after training with a stronger backbone on a larger dataset.The conclusion describes this as a new state-of-the-art result and reports that DINO outperforms previous ResNet-50-based models in 12- and 36-epoch settings.
  • Conclusion: DN-query training can produce duplicate boxes for one person, whereas the proposed method is shown alongside it for comparison.Fig. 8 displays only person boxes and highlights three duplicate boxes in the DN-query result.
  • Test Time Augmentations: Test-time augmentation does not provide an inspiring gain for DINO compared with traditional detectors.The exploration uses the large SwinL model and is included to compare with traditional detection models.

B Training Efficiency

The paper reports training resource measurements for DINO and compares architectural depth settings. These experiments frame DINO as both effective and efficient to train.

  • Training Resource Usage: DINO’s base-model GPU memory and training time are reported using 8 Nvidia A100 GPUs with a ResNet-50 backbone.The authors state that the results demonstrate training effectiveness and efficiency.
  • Training Resource Usage: Table 5 compares training efficiency across models with ResNet-50 under a common 8-GPU setup.The table includes training-efficiency measurements for different models and identifies framework-specific testing for Faster R-CNN.
  • Architecture Ablation: Table 6 ablates the numbers of encoder and decoder layers using ResNet-50 on COCO val2017 in the 12-epoch setting.The ablation uses 100 denoising queries without negative samples.

C Additional Analysis on our Model Components

The ablations examine decoder depth and denoising-query count, showing that both choices affect performance, with diminishing returns beyond 100 denoising queries.

  • Number of Encoder and Decoder Layers: A 3.0 AP drop results when decoder layers decrease from 6 to 2 with 6 encoder layers fixed.The analysis attributes this sensitivity to dynamic box updates and refinement across decoder layers.
  • Query Denoising: Increasing denoising queries improves performance significantly below 100 queries.The experiment uses the optimized dynamic denoising-group setting.
  • Query Denoising: Beyond 100 denoising queries, additional queries provide only small or sometimes worse performance improvements.Table 7 reports the ablation on COCO validation with a ResNet-50 backbone.

D.1 Dynamic DN groups

The dynamic DN-group procedure improves denoising efficiency by using multiple noised versions of ground-truth groups while accounting for variable object counts across minibatches.

  • Ground-Truth Grouping: DN-DETR forms one denoising ground-truth group from all labeled objects and boxes in each image.Multiple noised versions of this group are used during training.
  • Group Count: DN-DETR sets the number of denoising groups to five or ten according to model size.The method uses multiple noised versions of the ground-truth group during training.
  • Batching: Minibatch padding makes each image use the largest denoising-query count present in the batch.This padding is needed because DETR-like models train with minibatches and images contain different numbers of objects.

D.2 Large-Scale Model Pre-trianing

DINO is pre-trained on the large-scale Objects365 dataset before fine-tuning on COCO, using separate training schedules and GPU configurations.

  • Objects365 Pre-training: Objects365 provides over 1.7M annotated training images and 80,000 annotated validation images.The authors use the first 5,000 validation images for validation and add the remainder to training.
  • Pre-training Setup: DINO is pre-trained on Objects365 for 26 epochs using 64 Nvidia A100 GPUs.Each GPU uses a local batch size of one image.
  • COCO Fine-tuning: The model is fine-tuned on COCO for 18 epochs using 16 Nvidia A100 GPUs.The fine-tuning configuration also uses a local batch size of one image per GPU.

D.3 Other Implementation Details

The implementation combines standard Transformer, optimization, augmentation, multi-scale, loss, and query settings inherited from or aligned with prior DETR-like models.

  • Basic Hyper-parameters: The default model uses 6-layer encoder and decoder Transformers with a hidden feature dimension of 256.The initial learning rate is 1 × 10^-4.
  • Basic Hyper-parameters: The learning rate drops at epochs 11, 20, and 30 for the 12-, 24-, and 36-epoch ResNet-50 schedules.Each drop multiplies the learning rate by 0.1.
  • Optimization and Queries: DINO uses AdamW with weight decay 1×10^-4 and trains with batch size 16 on Nvidia A100 GPUs.The implementation uses 900 decoder queries to match the computation cost of 300 queries with 3 patterns.
  • Loss Function: Box regression uses L1 and GIOU losses, while classification uses focal loss with α = 0.25 and γ = 2.Auxiliary losses are added after each decoder layer, with intermediate losses after query selection.
  • Detailed Model Components: The pipeline adopts multi-scale features and deformable attention, while dynamic denoising groups improve denoising efficiency and reduce memory overhead.Prediction heads use unshared parameters across decoder layers.
  • Training Augmentation: Training uses random crop and scale augmentation with shorter image sides between 480 and 800 pixels and longer sides up to 1333 pixels.SwinL fine-tuning uses a 1.5× larger scale, with shorter sides between 720 and 1200 pixels and longer sides up to 2000 pixels.
  • Multi-scale Setting: Four-scale models use backbone stages 2–4 plus a downsampled stage-4 feature, while five-scale models additionally use stage 1.The configuration uses λ1 = 1.0, λ2 = 2.0, and 100 CDN pairs.

D.4 Detailed Hyper-parameters

This section points readers to Table 8 for the hyper-parameters used in the models, supporting reproduction of the reported results.

  • The hyper-parameter details are provided for readers who want to reproduce the results.
  • The section directs readers to Table 8 as the location of these implementation details.
  • Table 8 lists the hyper-parameters used in the models.
Loading 2203.03605v4…