Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Instances as Queries (2105.01928v3)

Published 5 May 2021 in cs.CV

Abstract: Recently, query based object detection frameworks achieve comparable performance with previous state-of-the-art object detectors. However, how to fully leverage such frameworks to perform instance segmentation remains an open problem. In this paper, we present QueryInst (Instances as Queries), a query based instance segmentation method driven by parallel supervision on dynamic mask heads. The key insight of QueryInst is to leverage the intrinsic one-to-one correspondence in object queries across different stages, as well as one-to-one correspondence between mask RoI features and object queries in the same stage. This approach eliminates the explicit multi-stage mask head connection and the proposal distribution inconsistency issues inherent in non-query based multi-stage instance segmentation methods. We conduct extensive experiments on three challenging benchmarks, i.e., COCO, CityScapes, and YouTube-VIS to evaluate the effectiveness of QueryInst in instance segmentation and video instance segmentation (VIS) task. Specifically, using ResNet-101-FPN backbone, QueryInst obtains 48.1 box AP and 42.8 mask AP on COCO test-dev, which is 2 points higher than HTC in terms of both box AP and mask AP, while runs 2.4 times faster. For video instance segmentation, QueryInst achieves the best performance among all online VIS approaches and strikes a decent speed-accuracy trade-off. Code is available at \url{https://github.com/hustvl/QueryInst}.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (8)
  1. Yuxin Fang (15 papers)
  2. Shusheng Yang (16 papers)
  3. Xinggang Wang (163 papers)
  4. Yu Li (378 papers)
  5. Chen Fang (157 papers)
  6. Ying Shan (252 papers)
  7. Bin Feng (44 papers)
  8. Wenyu Liu (146 papers)
Citations (236)

Summary

Instances as Queries: A Query-Based Instance Segmentation Approach

The paper "Instances as Queries" presents QueryInst: a novel query-based instance segmentation framework aiming to leverage parallel supervision on dynamic mask heads. The method builds upon recent advances in query-based object detection frameworks, such as DETR and Sparse R-CNN, reformulating instance segmentation to exploit the intrinsic one-to-one correspondence in object queries. This work explores the decoupling of the multi-stage proposal generation and segmentation tasks, proposing a coherent approach to enhance efficiency and performance.

Key Contributions

  1. Parallel Dynamic Mask Heads: QueryInst introduces dynamic mask heads featuring parallel supervision. This design utilizes one-to-one correspondence across object queries at different stages to refine mask RoI features adaptively. The main advantage is the bypassing of explicit multi-stage mask head connections, mitigating proposal distribution inconsistencies seen in traditional methods.
  2. Integration with Query-Based Detectors: The method synergizes with query-based object detection architectures by embedding dynamic convolutions within Mask RCNN-type architectures, allowing mask generation informed by the refined query representations. This architecture improves both joint detection and segmentation tasks through shared queries, enhancing cross-task communication.
  3. Superior Speed and Accuracy: The proposed method achieves significant improvements in speed and accuracy over state-of-the-art models like HTC. Empirically, using a ResNet-101-FPN backbone, QueryInst reported a 48.1 box AP and 42.8 mask AP on the COCO test-dev, surpassing HTC by 2 AP points in both categories while operating 2.4 times faster.
  4. Extensions to Video Instance Segmentation: QueryInst is extended to video instance segmentation (VIS) tasks by incorporating a simple tracking mechanism. Evaluation on YouTube-VIS highlights its ability to balance speed and accuracy, outperforming other online VIS approaches.

Numerical Results

  • COCO Benchmark: QueryInst with ResNet-101-FPN achieved 48.1 box AP and 42.8 mask AP, outperforming HTC by 2 points in both measures.
  • Cityscapes Benchmark: The method achieved 39.4 AP on the Cityscapes val set, offering superior performance compared to competitors.
  • VIS Benchmark: On YouTube-VIS, QueryInst-VIS outperformed several existing approaches, demonstrating its applicability to dynamic environments.

Theoretical and Practical Implications

The theoretical implications are significant in that QueryInst effectively bridges a gap between query-based object detection methods and instance segmentation tasks, an area with room for further exploration. The practical impacts are demonstrated in its ability to offer a faster and more efficient instance segmentation approach, applicable to both static and dynamic environments.

Future Directions

Potential future work could explore optimizing QueryInst for more complex scenarios or extending the idea to other segmentation tasks such as panoptic segmentation. Investigating the application of this framework to unsupervised or semi-supervised learning settings might also uncover new capabilities. The adaptability shown in QueryInst suggests promising pathways not only in computer vision but potentially in natural language processing, where similar query-based models are receiving increased attention.

In summary, QueryInst represents a robust and high-performing instance segmentation solution that advances the capabilities of query-based frameworks. Its innovations in architectural design and efficiency mark a substantial contribution to contemporary computer vision methodologies.

Github Logo Streamline Icon: https://streamlinehq.com
X Twitter Logo Streamline Icon: https://streamlinehq.com
Youtube Logo Streamline Icon: https://streamlinehq.com