Deep Layer Aggregation

By Fisher Yu, Dequan Wang, Trevor Darrell. 2017

1,537 citations recorded by the source corpus.

Visual recognition requires rich representations that span levels from low to high, scales from small to large, and resolutions from fine to coarse. Even with the depth of features in a convolutional network, a layer in isolation is not enough: compounding and aggregating these representations improves inference of what and where. Architectural efforts are exploring many dimensions for network backbones, designing deeper or wider architectures, but how to best aggregate layers and blocks across a network deserves further attention. Although skip connections have been incorporated to combine layers, these connections have been "shallow" themselves, and only fuse by simple, one-step operations. We augment standard architectures with deeper aggregation to better fuse information across layers. Our deep layer aggregation structures iteratively and hierarchically merge the feature hierarchy to make networks with better accuracy and fewer parameters. Experiments across architectures and tasks show that deep layer aggregation improves recognition and resolution compared to existing branching and merging schemes.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Model Architectures

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Model Architectures

Citations in the corpus

Objects as Points (2019, methodology): “(c): The original DLA-34 [58] for semantic segmentation.”
Ego4D: Around the World in 3,000 Hours of Egocentric Video (2021, background): “[233] Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell.”
AutoAugment: Learning Augmentation Policies from Data (2018, methodology): “Previous state-of-the-art ﬁne-tuned pre-trained weights on ImageNet and used deep layer aggregation to attain a 5.9% error rate [72].”
NAS-FPN: Learning Scalable Feature Pyramid Architecture for Object Detection (2019, background): “Many recent works [7, 16, 17, 34, 38, 39, 40, 43, 41] propose various crossscale connections or operations to combine features to gen-”
EPicker is an exemplar-based continual learning approach for knowledge accumulation in cryoEM particle picking (2022, methodology): “Several feature extraction networks are available for CenterNet, such as ResNet 35 and DLA 36 .”
BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learning (2018, methodology): “We provide image tagging classiﬁcation results using DLA-34 [37] in Figure 4.”
SqueezeNext: Hardware-Aware Neural Network Design (2018, methodology): “One such possibility is to use a tree structure instead of the typical feed forward architecture as proposed in [28].”
Large Scale Fine-Grained Categorization and Domain-Specific Transfer Learning (2018, neutral): “9 DLA [65] 85.”
Res2Net: A New Multi-Scale Backbone Architecture (2019, methodology): “For experiments on the ImageNet [34] dataset, due to our limited computational resources, we mainly use the ResNet-50 [17], ResNeXt-50 [43] and DLA-60 [47] as our baseline models.”
CLRNet: Cross Layer Refinement Network for Lane Detection (2022, methodology): “We adopt the ResNet [6] and DLA [30] as our pre-trained backbones.”
PPDM: Parallel Point Detection and Matching for Real-Time Human-Object Interaction Detection (2019, methodology): “The DLA-34 is a lightweight backbone network, and we apply a reﬁned version proposed in [35].”
A Real-Time Cross-Modality Correlation Filtering Method for Referring Expression Comprehension (2019, methodology): “The image encoder E(·) adopts the Deep Layer Aggregation (DLA) [31] architecture with deformable convolution [4].”
Deep Mixture of Experts via Shallow Embedding (2018, methodology): “To train DeepMoE we follow common training practices [11, 31].”
Track to Detect and Segment: An Online Multi-Object Tracker (2021, methodology): “Implementation Details: We adopt the same experimental settings as CenterTrack [63], such as backbone, image sizes, pretraining, score thresholds, etc. Speciﬁcally, we adopt the DLA-34 [60] as the backbone network φ ( · ) .”
Crossover Learning for Fast Online Video Instance Segmentation (2021, methodology): “Similar to the setup of [68, 6], we initialize CrossVIS with corresponding CondInst instance segmentation model [52, 26, 70, 34] pre-trained on COCO train2017 [36] with 1× schedule.”
Half-Real Half-Fake Distillation for Class-Incremental Semantic Segmentation (2021, positive): “…state-of-the-art semantic segmentation approaches based on the fully convolutional network (FCN) [23] have made remarkable progress in several ways, e.g. by modeling context information [50, 7, 49, 43, 19], re-covering the spatial details [8, 36, 20] or designing stronger networks [46, 42, 34].”
Dense Label Encoding for Boundary Discontinuity Free Rotation Detection (2020, background): “S in g le -s ta g e m et h o d s PIoU [4] DLA-34 [53] 80.”