Information-theoretic regularization for Multi-source Domain Adaptation

Park, Geon Yeong; Lee, Sang Wan

doi:10.1109/iccv48922.2021.00908

Cited by 20 publications

(5 citation statements)

References 21 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We use average precision (AP) and mean AP (mAP) as the evaluation metrics. In Table 1, we first compare the performances between our weighted MBD and MBCE [7] using the SWDA learning scheme [2] for simplicity. Because the original MBCE was designed for the classification, the detection performance decreased at both cross camera and cross time adaptations compared with SWDA.…”

Section: Resultsmentioning

confidence: 99%

“…The discriminator, which plays the most pivotal role in aligning domains, is designed for both high and low features [2]. Following MBCE [7], each discriminator makes the conditional probability of belonging to each domain as a value of 0 to 1. The equation of the MBCE can be simply described as follows:…”

Section: Weighted Multiple Binary Discriminator (Mbd)mentioning

confidence: 99%

$$\begin{equation} P\left( x \right) = \left\{ \def\eqcellsep{&}\begin{array}{ll} {\log {p_i}\left( x \right),}&{{\rm{if\;}}x \in i{\rm{th\;Domain}},}\\[6pt] {\log \left( {1 - {p_i}\left( x \right)} \right),}&{{\rm{otherwise}},} \end{array} \right.

…”

Section: Weighted Multiple Binary Discriminator (Mbd)mentioning

confidence: 99%

“…From the viewpoint of the classification task, there are some efforts in the MSDA classification [5, 6] but most MSDA methods have attempted to overcome the multiple domain discrepancy by using multiple domain discriminators. Recently, in [7], they proposed the multiple binary domain discriminator based on the Multiple Binary Cross Entropy (MBCE) where adversarial loss is minimized over multiple domains with a single‐domain discriminator in the classification task. Going back to the object detection, there are fewer studies on MSDA for detection than for classification.…”

Section: Introductionmentioning

confidence: 99%

“…Unlike [8,9], we first make the efficient single networkbased backbone with a single ROI head for the fast object detection. Directly inspired from [7], we propose a new multiple binary domain discriminator that makes use of the different weights depending on how large the difference between the domains is for aligning the multiple domains for object detection. Note that, unlike classification, multiple different objects in object detection can exist in a single image or the background region is wider than the objects to be detected.…”

mentioning

confidence: 99%

See 4 more Smart Citations

Weighted alignment‐based multi‐source domain adaptation for object detection

Han

Woo

Hwang

et al. 2023

Electronics Letters

View full text Add to dashboard Cite

Beyond single‐source domain adaption (DA) for object detection, multi‐source domain adaptation for object detection is another challenge because the authors should solve the multiple domain shifts between the source and target domains as well as between multiple source domains. In this letter, the authors propose a novel multi‐source domain adaptation via weighted alignment for object detection where the authors adopt a teacher–student framework. The authors first propose the weighted multiple binary discriminator (MBD) to align the multiple domains considering individual domain shifts. The authors also design the weighted class balance loss (CBL) that aligns the different weights to efficiently learn the detection model even if the number of objects is not balanced in an image under the teacher–student learning scheme. The authors empirically prove the superiority of our method on widely used benchmarks such as Cityscapes, KITTI, and BDD100k datasets.

show abstract

Section: Resultsmentioning

confidence: 99%

Section: Weighted Multiple Binary Discriminator (Mbd)mentioning

confidence: 99%