Probabilistic two-stage detection

By Xingyi Zhou, Vladlen Koltun, Philipp Krahenbuhl. 2021

272 citations recorded by the source corpus.

We develop a probabilistic interpretation of two-stage object detection. We show that this probabilistic interpretation motivates a number of common empirical training practices. It also suggests changes to two-stage detection pipelines. Specifically, the first stage should infer proper object-vs-background likelihoods, which should then inform the overall score of the detector. A standard region proposal network (RPN) cannot infer this likelihood sufficiently well, but many one-stage detectors can. We show how to build a probabilistic two-stage detector from any state-of-the-art one-stage detector. The resulting detectors are faster and more accurate than both their one- and two-stage precursors. Our detector achieves 56.4 mAP on COCO test-dev with single-scale testing, outperforming all published results. Using a lightweight backbone, our detector achieves 49.2 mAP on COCO at 33 fps on a Titan Xp, outperforming the popular YOLOv4 model.

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

1. Jianfeng Gao
2. Mohit Bansal
3. Dacheng Tao
4. Jianwei Yang
5. Trevor Darrell
6. Lu Yuan
7. Quoc V. Le
8. Armand Joulin

In Computer Vision and Pattern Recognition

1. Luc Van Gool
2. Dacheng Tao
3. Ming-Hsuan Yang
4. Xiaogang Wang
5. Trevor Darrell
6. Jitendra Malik
7. Cordelia Schmid
8. Andrew Zisserman

Citations in the corpus

• RegionCLIP: Region-based Language-Image Pretraining (2021, methodology): “Inspired by [47, 64], we fused RPN objectness scores and category confidence scores by geometry mean.”
• Exploring Plain Vision Transformer Backbones for Object Detection (2022, methodology): “We follow the same model and training details as used for the COCO systemlevel comparison plus two common LVIS practices: we use the federated loss from [57] and sample images with repeat factor sampling [22].”
• Simple Open-Vocabulary Object Detection with Vision Transformers (2022, methodology): “Finally, we randomly sample pseudonegative labels for each image until there are at least 50 negative labels [47].”
• EVA-02: A Visual Representation for Neon Genesis (2023, methodology): “The model architecture as well as the hyper-parameters for COCO and LVIS are almost the same, except we use federated loss [149] and repeat factor sampling [50] following ViTDet on LVIS.”
• Detecting Twenty-thousand Classes using Image-level Supervision (2022, methodology): “Federated Loss [70] subsamples classes per-iteration to mimic the federated annotation [16].”
• You Only Learn One Representation: Unified Network for Multiple Tasks (2021, unknown): “55.5% 73.4% 60.8% 16 EfficientDet [13] 55.1% 74.3% 59.9% 6.5 SwinTransformer [10] 57.7% – – – CenterNet2 [26] 56.4% 74.0% 61.6% – CopyPaste [6] 57.3% – – – • YOLOR-P6 has same architecture as YOLOv4-P6-light, we replace all Mish activation in YOLOv4-P6-light by SiLU activation.”
• DETRs with Collaborative Hybrid Assignments Training (2022, methodology): “In particular, we use the same Co-DINO-Deformable-DETR++ as the model on COCO but choose FedLoss [42] as the classification loss to remedy the impact of unbalanced data distribution.”
• Video-ChatGPT: Towards Detailed Video Understanding via Large Vision and Language Models (2023, methodology): “For GRIT [27], we use ViT-B version with CenterNet2 [31].”
• Multimodal Virtual Point 3D Detection (2021, methodology): “Figure 2: Comparison between state-of-the-art image-based 2D detector [73] and point cloud based 3D detector [66].”
• EVA: Exploring the Limits of Masked Visual Representation Learning at Scale (2022, methodology): “0), expect we use federated loss [122] and repeat factor sampling [37] following ViTDet on LVISv1.”
• TPH-YOLOv5: Improved YOLOv5 Based on Transformer Prediction Head for Object Detection on Drone-captured Scenarios (2021, unknown): “2) two-stage detectors: VFNet [59], CenterNet2 [62].”
• Instances as Queries (2021, background): “For each stage, the number and distribution of training samples are quite different, and there is no explicit and intrinsic correspondence for each individual proposal across different stages [59].”
• DiffusionDet: Diffusion Model for Object Detection (2022, methodology): “Following [111], we choose | S | = 50 in all experiments.”