SimROD: A Simple Adaptation Method for Robust Object Detection

Ramamonjison, Rindra; Banitalebi-Dehkordi, Amin; Kang, Xinyu; Bai, Xiaolong; Zhang, Yong

doi:10.1109/iccv48922.2021.00355

Cited by 29 publications

(14 citation statements)

References 20 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…2) Fair comparison. The fairness and agreement of the benchmark comparison have been proven in recently published literature [4,15,24,26,36] for single-stage object detectors due to the comparable source only results and adaptation gains. Besides, we also report the fair adaptation gains in benchmark comparison to demonstrate our effectiveness in terms of domain adaptation.…”

Section: B1 Baseline Selectionmentioning

confidence: 59%

“…Reasons for the singe-stage baseline. In this paper, we mainly focus on the domain adaptation for singe-stage object detectors as lots of recently published works [4,15,21,24,26,36], and we select the single-stage detector as the baseline because of the following two main reasons. 1) Discarding RPN.…”

Section: B1 Baseline Selectionmentioning

confidence: 99%

“…Moreover. most of the latest adaptation works [4,15,24,26,36] are based on the single-stage detectors [27,37], and we aim to present a comparison with them using same baseline model. Potentials for the two-stage extension.…”

Section: B1 Baseline Selectionmentioning

confidence: 99%

See 2 more Smart Citations

SIGMA: Semantic-complete Graph Matching for Domain Adaptive Object Detection

Li¹,

Liu²,

Yuan³

2022

Preprint

View full text Add to dashboard Cite

Domain Adaptive Object Detection (DAOD) leverages a labeled domain to learn an object detector generalizing to a novel domain free of annotations. Recent advances align class-conditional distributions by narrowing down crossdomain prototypes (class centers). Though great success, they ignore the significant within-class variance and the domain-mismatched semantics within the training batch, leading to a sub-optimal adaptation. To overcome these challenges, we propose a novel SemantIc-complete Graph MAtching (SIGMA) framework for DAOD, which completes mismatched semantics and reformulates the adaptation with graph matching. Specifically, we design a Graph-embedded Semantic Completion module (GSC) that completes mismatched semantics through generating hallucination graph nodes in missing categories. Then, we establish crossimage graphs to model class-conditional distributions and learn a graph-guided memory bank for better semantic completion in turn. After representing the source and target data as graphs, we reformulate the adaptation as a graph matching problem, i.e., finding well-matched node pairs across graphs to reduce the domain gap, which is solved with a novel Bipartite Graph Matching adaptor (BGM). In a nutshell, we utilize graph nodes to establish semantic-aware node affinity and leverage graph edges as quadratic constraints in a structure-aware matching loss, achieving finegrained adaptation with a node-to-node graph matching. Extensive experiments verify that SIGMA outperforms existing works significantly. Our code is available at https: //github.com/CityU-AIM-Group/SIGMA.

show abstract

Section: B1 Baseline Selectionmentioning

confidence: 59%

Section: B1 Baseline Selectionmentioning

confidence: 99%

See 1 more Smart Citation

SIGMA: Semantic-complete Graph Matching for Domain Adaptive Object Detection

Li¹,

Liu²,

Yuan³

2022

Preprint

View full text Add to dashboard Cite

show abstract

“…To solve these problems, inspired by Ramamonjison et al ( 2021 ), we propose to sample one image from the source domain

and the latest target domain

, respectively, halve their long edges and assemble them from left to right in a stochastic order. We adopt similar ways to transform and concatenate corresponding data for annotations.…”

Section: Domain-incremental Adaptionmentioning

confidence: 99%

Brain-Inspired Domain-Incremental Adaptive Detection for Autonomous Driving

et al. 2022

View full text Add to dashboard Cite

Most existing methods for unsupervised domain adaptation (UDA) only involve two domains, i.e., source domain and the target domain. However, such trained adaptive models have poor performance when applied to a new domain without learning. Moreover, using UDA methods to adapt from the source domain to the new domains will lead to catastrophic forgetting of the previous target domain. To handle these issues, inspired by the ability to balance the maintenance of old knowledge and learning new knowledge of the human brain, in this article, we propose a new incremental learning framework for domain-incremental cases, which can harmonize the memorability and discriminability of the existing and the novel domains. By this means, the model can imitate the learning process of the human brain and, thus, improve its adaptability. To evaluate the effectiveness of the proposed methods, we conduct two groups of experiments, including virtual-to-real and diverse-weather cases. The experimental results demonstrate that our approach can avoid catastrophic forgetting, mitigate performance degradation in the previous domains, and improve the object detection accuracy of the novel target domain significantly.

show abstract

“…Consistency regularization is initially proposed in semisupervised image learning tasks [60]. Recently, this architecture and its advanced variants have achieved state-of-theart performance in the semi-supervised learning (SSL) [5,75,79,83,86] and unsupervised domain adaptation (UDA) benchmarks [3,13,17,34,55,76,87,88]. Their frameworks include a student and teacher model, where the teacher model uses the EMA weight of the student models.…”

Section: Consistency Regularizationmentioning

confidence: 99%

Domain Adaptive Semantic Segmentation via Regional Contrastive Consistency Regularization

Zhou,

Zhuang,

et al. 2021

Preprint

View full text Add to dashboard Cite

Unsupervised domain adaptation (UDA) aims to bridge the domain shift between the labeled source domain and the unlabeled target domain. However, most existing works perform the global-level feature alignment for semantic segmentation, while the local consistency between the regions has been largely neglected, and these methods are less robust to changing of outdoor environments. Motivated by the above facts, we propose a novel and fully end-to-end trainable approach, called regional contrastive consistency regularization (RCCR) for domain adaptive semantic segmentation. Our core idea is to pull the similar regional features extracted from the same location of different images to be closer, and meanwhile push the features from the different locations of the two images to be separated. We innovatively propose momentum projector heads, where the teacher projector is the exponential moving average of the student. Besides, we present a region-wise contrastive loss with two sampling strategies to realize effective regional consistency. Finally, a memory bank mechanism is designed to learn more robust and stable region-wise features under varying environments. Extensive experiments on two common UDA benchmarks, i.e., GTAV to Cityscapes and SYN-THIA to Cityscapes, demonstrate that our approach outperforms the state-of-the-art methods.

show abstract

SimROD: A Simple Adaptation Method for Robust Object Detection

Cited by 29 publications

References 20 publications

SIGMA: Semantic-complete Graph Matching for Domain Adaptive Object Detection

SIGMA: Semantic-complete Graph Matching for Domain Adaptive Object Detection

Brain-Inspired Domain-Incremental Adaptive Detection for Autonomous Driving

Domain Adaptive Semantic Segmentation via Regional Contrastive Consistency Regularization

Contact Info

Product

Resources

About