DetectoRS: Detecting Objects with Recursive Feature Pyramid and Switchable Atrous Convolution

By Siyuan Qiao, Liang-Chieh Chen, Alan L. Yuille. 2020

1,026 citations recorded by the source corpus.

Many modern object detectors demonstrate outstanding performances by using the mechanism of looking and thinking twice. In this paper, we explore this mechanism in the backbone design for object detection. At the macro level, we propose Recursive Feature Pyramid, which incorporates extra feedback connections from Feature Pyramid Networks into the bottom-up backbone layers. At the micro level, we propose Switchable Atrous Convolution, which convolves the features with different atrous rates and gathers the results using switch functions. Combining them results in DetectoRS, which significantly improves the performances of object detection. On COCO test-dev, DetectoRS achieves state-of-the-art 55.7% box AP for object detection, 48.5% mask AP for instance segmentation, and 50.0% PQ for panoptic segmentation. The code is made publicly available1.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

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

Citations in the corpus

Simple Copy-Paste is a Strong Data Augmentation Method for Instance Segmentation (2020, methodology): “We report state-of-the-art 2 results on the COCO benchmark surpassing SpineNet [11] by 2.8 AP and DetectoRS [43] by 0.6 AP.”
MaX-DeepLab: End-to-End Panoptic Segmentation with Mask Transformers (2020, methodology): “This result also outperforms the best box-based method DetectoRS [76] with TTA by 1.7% PQ, closing the large gap between box-based and box-free methods on COCO for the first time.”
Detecting tiny objects in aerial images: A normalized Wasserstein distance and a new benchmark (2022, positive): “In this paper, we select anchor-based object detectors (Ren et al., 2015; Cai and Vas., 2018; Lin et al., 2017b; Qiao et al., 2021; Zhang et al., 2020) which simultaneously hold the state-of-the-art performance on normal object detection tasks (Qiao et al.”
Transformer-Based Visual Segmentation: A Survey (2023, methodology): “Several works [83], [119] design stronger FPNs via cross-scale modeling using dilation or deformable convolution.”
Scaling Wide Residual Networks for Panoptic Segmentation (2020, methodology): “. Building on top of WR-41, we further incorporate the simpliﬁed Squeeze-and-Excitation (SE) module [36,49] (where only one fully connected layer is used), and the Switchable Atrous Convolution (SAC) [69], forming our baseline model. In Fig.2, we visualize the simpliﬁed SE module and SAC operation. ”
CoDeNet: Efficient Deployment of Input-Adaptive Object Detection on Embedded FPGAs (2021, methodology): “Many neural network designs with top accuracy [29, 40] for object detection on the COCO dataset [22] use deformable convolution in their design.”
Towards Large-Scale Small Object Detection: Survey and Benchmarks (2022, background): “This simple yet effective design has become an essential component in feature extractor and inspires a series of remarkable variants, e.g., NAS-FPN [72], and Recursive-FPN [73].”
YOLO-FaceV2: A Scale and Occlusion Aware Face Detector (2022, methodology): “Currently, the main method to solve the problem of varying scales is constructing a pyramid to fuse the multi-scale features of faces [17, 18, 19, 20].”
A Normalized Gaussian Wasserstein Distance for Tiny Object Detection (2021, positive): “After that, lots of methods are proposed to further improve FPN performance, including PANet [17], BiFPN [26], Recursive-FPN [20].”
CBNet: A Composite Backbone Network Architecture for Object Detection (2021, methodology): “There are several relevant detectors, such as DetectoRS [74] that composites both backbone and FPN and Joint-DetNAS [47] searches for the model scaling strategy.”
k-means Mask Transformer (2022, neutral): “0 CMT-DeepLab [106] Axial-R104-RFN† [80] 270M 1114G 3.”
CMT-DeepLab: Clustering Mask Transformers for Panoptic Segmentation (2022, methodology): “Instead of using the complicated fusion module proposed in [7], we simply average the features for fusion.”