Source-free domain adaptation (SFDA) aims to adapt models to unlabeled target domains without access to source data. Recent SFDA approaches leverage foundation models with strong generalization capabilities to adapt effectively for target domains. However, previous methods that align predictions from distinct models result in substantial performance degradation on the source domain. In this paper, we introduce a Generalized SFDA method that maintains source domain performance while the target model performs collaborative learning with the foundation model. Our method focuses on the initial embedding of the target model, which starts from the source model. Specifically, we constrain the change in feature space of the target data based on the confidence scores to prevent representation drift for the source domain. Simultaneously, we gradually decouple the foundation model from the target model, maintaining performance on the source domain while improving performance on the target domain. We evaluate our method using standard domain adaptation benchmarks, including Office–Home, DomainNet-126, and VisDA. Furthermore, by evaluating our method under the Generalized SFDA and Continual SFDA protocols, we demonstrate that it not only adapts effectively to the target domain but also preserves performance on the source domain.

Knowledge distillation transfers knowledge from a high-capacity teacher network to a compact student network, but a large capacity gap often limits the student's ability to fully benefit from the teacher's guidance. In semantic segmentation, another major challenge is the difficulty in predicting accurate object boundaries, as even strong teacher models can produce ambiguous or imprecise outputs. To address both challenges, we present Knowledge Tailoring, a novel distillation framework that adapts the teacher's knowledge to better match the student's representational capacity and learning dynamics. Much like a tailor adjusts an oversized suit to fit the wearer's shape, our method reshapes the teacher's abundant but misaligned knowledge into a form more suitable for the student. KT introduces feature tailoring, which restructures intermediate features based on channel-wise correlation to narrow the representation gap, and logit tailoring, which improves boundary prediction by refining class-specific logits. The tailoring strategy evolves throughout training, offering guidance that aligns with the student's progress. Experiments on Cityscapes, Pascal VOC, and ADE20K confirm that KT consistently enhances performance across a variety of architectures including DeepLabV3, PSPNet, and SegFormer. Our code is available for https://github.com/seok-hwa/KT.

Domain adaptation for object detection tasks typically focus on visible-spectrum domains. However, transferring knowledge from RGB to thermal infrared (TIR) images remains challenging due to significant domain discrepancies. To address this challenge, we propose SwitchMix, a dual-level mixedteacher framework designed to progressively narrow the crossmodality discrepancy for unsupervised thermal object detection. The proposed method integrates image-level and sample-level domain mixing to align modality specific spectral and semantic representations, while a teacher-student paradigm gradually shifts supervision from RGB to thermal domains. By periodically activating two complementary teachers, the model assimilates both local and global domain knowledge. Experiments on the FLIR and KAIST datasets demonstrate SwitchMix consistently outperforms existing state-of-the-art unsupervised domain adaptation approaches in both accuracy and robustness. These results highlight the potential of SwitchMix as a generalizable domain adaptation solution for applications such as autonomous driving, defense systems, and intelligent surveillance.

Source-Free Domain Adaptation (SFDA) aims to adapt a pre-trained source model to an unlabeled target domain without access to source data. Recent advances in Foundation Models (FMs) have introduced new opportunities for leveraging external semantic knowledge to guide SFDA. However, relying on a single FM is often insufficient, as it tends to bias adaptation toward a restricted semantic coverage, failing to capture diverse contextual cues under domain shift. To overcome this limitation, we propose a Collaborative Multi-foundation Adaptation (CoMA) framework that jointly leverages two different FMs (e.g., CLIP and BLIP) with complementary properties to capture both global semantics and local contextual cues. Specifically, we employ a bidirectional adaptation mechanism that (1) aligns different FMs with the target model for task adaptation while maintaining their semantic distinctiveness, and (2) transfers complementary knowledge from the FMs to the target model. To ensure stable adaptation under mini-batch training, we introduce Decomposed Mutual Information (DMI) that selectively enhances true dependencies while suppressing false dependencies arising from incomplete class coverage. Extensive experiments demonstrate that our method consistently outperforms existing state-of-the-art SFDA methods across four benchmarks, including Office-31, Office-Home, DomainNet-126, and VisDA, under the closed-set setting, while also achieving best results on partial-set and open-set variants.

Vision Foundation Model (VFM) such as the Segment Anything Model (SAM) and Contrastive Language-Image Pre-training Model (CLIP) has shown promising performance for segmentation and detection tasks. However, although SAM excels in fine-grained segmentation, it faces major challenges when applying it to semantic-aware segmentation. While CLIP exhibits a strong semantic understanding capability via aligning the global features of language and vision, it has deficiencies in fine-grained segmentation tasks. Human parsing requires to segment human bodies into constituent parts and involves both accurate fine-grained segmentation and high semantic understanding of each part. Based on traits of SAM and CLIP, we formulate high efficient modules to effectively integrate features of them to benefit human parsing. We propose a Semantic-Refinement Module to integrate semantic features of CLIP with SAM features to benefit parsing. Moreover, we formulate a high efficient Fine-tuning Module to adjust the pretrained SAM for human parsing that needs high semantic information and simultaneously demands spatial details, which significantly reduces the training time compared with full-time training and achieves notable performance. Extensive experiments demonstrate the effectiveness of our method on LIP, PPP, and CIHP databases.

Continual learning aims to refine model parameters for new tasks while retaining knowledge from previous tasks. Recently, prompt-based learning has emerged to leverage pre-trained models to be prompted to learn subsequent tasks without the reliance on the rehearsal buffer. Although this approach has demonstrated outstanding results, existing methods depend on preceding task-selection process to choose appropriate prompts. However, imperfectness in task-selection may lead to negative impacts on the performance particularly in the scenarios where the number of tasks is large or task distributions are imbalanced. To address this issue, we introduce I-Prompt, a task-agnostic approach focuses on the visual semantic information of image tokens to eliminate task prediction. Our method consists of semantic prompt matching, which determines prompts based on similarities between tokens, and image token-level prompting, which applies prompts directly to image tokens in the intermediate layers. Consequently, our method achieves competitive performance on four benchmarks while significantly reducing training time compared to state-of-the-art methods. Moreover, we demonstrate the superiority of our method across various scenarios through extensive experiments.

Test-time adaptation paradigm provides flexibility towards domain shifts by performing immediate adaptation on unlabeled target data from the source model. Vision-Language Models (VLMs) leverage their generalization capabilities for diverse downstream tasks, and test-time prompt tuning has emerged as a prominent solution for adapting VLMs. In this work, we explore contrastive VLMs and identify the modality gap caused by a single dominant feature dimension across modalities. We observe that the dominant dimensions in both text and image modalities exhibit high predictive sensitivity, and that constraining their influence can improve calibration error. Building on this insight, we propose dimensional entropy maximization that regularizes the distribution of textual features toward uniformity to mitigate the dependency of dominant dimensions. Our method alleviates the degradation of calibration performance in test-time prompt tuning, offering a simple yet effective solution to enhance the reliability of VLMs in real-world deployment scenarios.

Test-time adaptation aims to adapt to realistic environments in an online manner by learning during test time. Entropy minimization has emerged as a principal strategy for test-time adaptation due to its efficiency and adaptability. Nevertheless, it remains underexplored in continual test-time adaptation, where stability is more important. We observe that the entropy minimization method often suffers from model collapse, where the model converges to predicting a single class for all images due to a trivial solution. We propose ranked entropy minimization to mitigate the stability problem of the entropy minimization method and extend its applicability to continuous scenarios. Our approach explicitly structures the prediction difficulty through a progressive masking strategy. Specifically, it gradually aligns the model's probability distributions across different levels of prediction difficulty while preserving the rank order of entropy. The proposed method is extensively evaluated across various benchmarks, demonstrating its effectiveness through empirical results. Our code is available at https://github.com/pilsHan/rem

Training on test-time data enables deep learning models to adapt to dynamic environmental changes, enhancing their practical applicability. Online adaptation from source to target domains is promising but it remains highly reliant on the performance of source pretrained model. In this paper, we investigate whether test-time adaptation (TTA) methods can continuously improve models trained via self-supervised learning (SSL) without relying on source pretraining. We introduce a self-supervised TTA protocol after observing that existing TTA approaches struggle when directly applied to self-supervised models with low accuracy on the source domain. Furthermore, we propose a collaborative learning framework that integrates SSL and TTA models, leveraging contrastive learning and knowledge distillation for stepwise representation refinement. We validate our method on diverse self-supervised models, including DINO, MoCo, and iBOT, across TTA benchmarks. Extensive experiments validate the effectiveness of our approach in SSL, showing that it achieves competitive performance even without source pretraining.

Unsupervised domain adaptation (UDA) methods for learning domain invariant representations have achieved remarkable progress. However, most of the studies were based on direct adaptation from the source domain to the target domain and have suffered from large domain discrepancies. In this paper, we propose a UDA method that effectively handles such large domain discrepancies. We introduce a fixed ratio-based mixup to augment multiple intermediate domains between the source and target domain. From the augmented-domains, we train the source-dominant model and the target-dominant model that have complementary characteristics. Using our confidence-based learning methodologies, e.g., bidirectional matching with high-confidence predictions and self-penalization using low-confidence predictions, the models can learn from each other or from its own results. Through our proposed methods, the models gradually transfer domain knowledge from the source to the target domain. Extensive experiments demonstrate the superiority of our proposed method on three public benchmarks: Office-31, Office-Home, and VisDA-2017.(1)

Unsupervised domain adaptation for object detection (UDAOD) aims to reduce the gap between the labeled source domain and the unlabeled target domain. In the driving scenes, there are distinct unique characteristics that differentiate between the objects, both spatially and categorically. These properties largely maintain their invariance across domains, enabling the effective training of object detectors in the target domain. To consider this, we introduce the domain-invariant object concentration framework, which combines instance-level and image-level approaches to efficiently utilizing domain-invariant object knowledge. At the instance-level, we propose a target surrogate selection module. This module leverages the unique information of object categories to match regions of interest (ROIs) between both domains, thereby enabling effective training on the unlabeled target domain. At the image level, we introduce PutMix, which utilizes domain-invariant knowledge of common object positions in driving scenes based on our statistically defined object crowded area. To validate our method, we experiment across four driving scenarios using four different datasets.

Unsupervised domain adaptation aims to mitigate the domain gap between the source and the target domains. Despite domain shifts, we have observed intrinsic knowledge that spans across domains for object detection in urban driving scenes. First, it includes consistent characteristics of objects within the same category of extracted ROIs. Second, it encompasses the similarity of patterns within the extracted ROIs, relating to the positions of the foreground and background during object detection. To utilize these, we present DuPDA, a method that effectively adapts object detectors to target domains by leveraging domain-invariant knowledge to separable objectness for training. Specifically, we construct categorical and regional prototypes, each of which operates through their specialized moving alignments. These prototypes serve as valuable references for training unlabeled target objects using similarity. Leveraging these prototypes, we determine and utilize a boundary that trains separately the foreground and background regions within the target ROIs, thereby transferring the knowledge to focus on each respective region. Our DuPDA surpasses previous state-of-the-art methods in various evaluation protocols on six benchmarks.

Human intelligence gradually accepts new information and accumulates knowledge throughout the lifespan. However, deep learning models suffer from a catastrophic forgetting phenomenon, where they forget previous knowledge when acquiring new information. Class-Incremental Learning aims to create an integrated model that balances plasticity and stability to overcome this challenge. In this paper, we propose a selective regularization method that accepts new knowledge while maintaining previous knowledge. We first introduce an asymmetric feature distillation method for old and new classes inspired by cognitive science, using the gradient of classification and knowledge distillation losses to determine whether to perform pattern completion or pattern separation. We also propose a method to selectively interpolate the weight of the previous model for a balance between stability and plasticity, and we adjust whether to transfer through model confidence to ensure the performance of the previous class and enable exploratory learning. We validate the effectiveness of the proposed method, which surpasses the performance of existing methods through extensive experimental protocols using CIFAR-100, ImageNet-Subset, and ImageNet-Full.

Domain adaptation for object detection typically entails transferring knowledge from one visible domain to another visible domain. However, there are limited studies on adapting from the visible to the thermal domain, because the domain gap between the visible and thermal domains is much larger than expected, and traditional domain adaptation can not successfully facilitate learning in this situation. To overcome this challenge, we propose a Distinctive Dual-Domain Teacher (D3T) framework that employs distinct training paradigms for each domain. Specifically, we segregate the source and target training sets for building dual-teachers and successively deploy exponential moving average to the student model to individual teachers of each domain. The framework further incorporates a zigzag learning method between dual teachers, facilitating a gradual transition from the visible to thermal domains during training. We validate the superiority of our method through newly designed experimental protocols with well-known thermal datasets, i.e., FLIR and KAIST. Source code is available at https://github.com/EdwardDo69/D3T .

Multi-target domain adaptation (MTDA) for semantic segmentation poses a significant challenge, as it involves multiple target domains with varying distributions. The goal of MTDA is to minimize the domain discrepancies among a single source and multi-target domains, aiming to train a single model that excels across all target domains. Previous MTDA approaches typically employ multiple teacher architectures, where each teacher specializes in one target domain to simplify the task. However, these architectures hinder the student model from fully assimilating comprehensive knowledge from all target-specific teachers and escalate training costs with increasing target domains. In this paper, we propose an ouroboric domain bridging (OurDB) framework, offering an efficient solution to the MTDA problem using a single teacher architecture. This framework dynamically cycles through multiple target domains, aligning each domain individually to restrain the biased alignment problem, and utilizes Fisher information to minimize the forgetting of knowledge from previous target domains. We also propose a context-guided class-wise mixup (CGMix) that leverages contextual information tailored to diverse target contexts in MTDA. Experimental evaluations conducted on four urban driving datasets (i.e., GTA5, Cityscapes, IDD, and Mapillary) demonstrate the superiority of our method over existing state-of-the-art approaches.

The dominant backbones of neural networks for scene parsing consist of multiple stages, where feature maps in different stages often contain varying levels of spatial and semantic information. High-level features convey more semantics and fewer spatial details, while low-level features possess fewer semantics and more spatial details. Consequently, there are semantic-spatial gaps among features at different levels, particularly in human parsing tasks. Many existing approaches directly upsample multi-stage features and aggregate them through addition or concatenation, without addressing the semantic-spatial gaps present among these features. This inevitably leads to spatial misalignment, semantic mismatch, and ultimately misclassification in parsing, especially for human parsing that demands more semantic information and more fine details of feature maps for the reason of intricate textures, diverse clothing styles, and heavy scale variability across different human parts. In this paper, we effectively alleviate the long-standing challenge of addressing semantic-spatial gaps between features from different stages by innovatively utilizing the subtraction and addition operations to recognize the semantic and spatial differences and compensate for them. Based on these principles, we propose the Channel and Spatial Enhancement Network (CSENet) for parsing, offering a straightforward and intuitive solution for addressing semantic-spatial gaps via injecting high-semantic information to lower-stage features and vice versa, introducing fine details to higher-stage features. Extensive experiments on three dense prediction tasks have demonstrated the efficacy of our method. Specifically, our method achieves the best performance on the LIP and CIHP datasets and we also verify the generality of our method on the ADE20K dataset.

Point-cloud based 3D object detectors recently have achieved remarkable progress. However, most studies are limited to the development of network architectures for improving only their accuracy without consideration of the computational efficiency. In this paper, we first propose an autoencoder-style framework comprising channel-wise compression and decompression via interchange transfer-based knowledge distillation. To learn the map-view feature of a teacher network, the features from teacher and student networks are independently passed through the shared autoencoder; here, we use a compressed representation loss that binds the channel-wised compression knowledge from both student and teacher networks as a kind of regularization. The decompressed features are transferred in opposite directions to reduce the gap in the interchange reconstructions. Lastly, we present an head attention loss to match the 3D object detection information drawn by the multi-head self-attention mechanism. Through extensive experiments, we verify that our method can train the lightweight model that is well-aligned with the 3D point cloud detection task and we demonstrate its superiority using the well-known public datasets; e.g., Waymo and nuScenes.

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.

In computer vision, human parsing is challenging due to its demand for accurate human region location and semantic partitioning. This dense prediction task needs powerful computation and high-precision models. To enable real-time parsing on resource-limited devices, the authors introduced a lightweight model using ResNet18 as a core network . The authors simplified the pyramid module, improving context clarity and reducing complexity. The authors integrated a spatial attention fusion strategy to counter precision loss in the light-weighting process. Traditional models, despite their segmentation precision, are limited by their computational complexity and extensive parameters. The authors implemented knowledge distillation (KD) techniques to enhance the authors' lightweight network's accuracy. Traditional methods can fail to learn useful knowledge with significant network differences. Hence, the authors used a novel distillation approach based on inter-class and intra-class relations in prediction outcomes, noticeably improving parsing accuracy. The authors' experiments on the Look into Person (LIP) dataset show that their lightweight model significantly reduces parameters while maintaining parsing precision and enhancing inference speed.