Unsupervised Domain Adaptation (UDA) is essential for adapting object detection systems to diverse operational environments without requiring domain-specific labeled data. In parallel, real-time performance is crucial for deploying these detectors in intelligent vehicles. This paper introduces Uncertainty-Guided Adaptive Knowledge Distillation (UGAKD), a novel framework designed to enhance UDA while simultaneously reducing model size through targeted knowledge distillation. Given that adversarial learning is a common approach in UDA, often utilizing a domain classifier to identify domain-invariant features, UGAKD leverages the localization of these domain-invariant features to guide the distillation process. Furthermore, we propose a two-stage, difficulty-aware training scheme to facilitate learning, which emphasizes domain-invariant features to boost distillation efficacy. Experimental results across several challenging scenarios, including transitions from synthetic to real-world environments, varying weather conditions, and shifts between real and stylized domains, demonstrate that UGAKD effectively reduces model complexity while improving detection accuracy. Specifically, UGAKD decreases the number of parameters by over 37% and FLOPs by over 47%. Compared with the baseline fine-grained feature imitation method, UGAKD achieves an mAP improvement of 1.0% to 1.7%, highlighting its effectiveness in maintaining robust object detection across diverse settings and its suitability for applications that require both efficiency and adaptability.
Wei-Lun Tseng, Yi-Lun Wu, Yung-Hui Li et al.· ACM Transactions on Intellig...· 0 citations
Text-to-image (T2I) diffusion models have achieved striking progress but still struggle to synthesize rare concepts involving unusual attribute-object pairings, often resulting in concept omission or semantic drift where a dominant entity overwhelms the generation. Tracing these failures to a lack of compositional balance during the denoising trajectory, we propose RADIANCE, a training-free framework that treats inference as a closed-loop feedback process. RADIANCE augments pretrained backbones with three modular components: (1) a Compositional Similarity Monitor (CSM) that tracks the emergence of objects and attributes in intermediate latents via CLIP-based feedback; (2) a Bidirectional Scale Controller (BSC) that applies a reactive"restoring force"using positive and negative IP-Adapter scales to rebalance biased trajectories; and (3) a Feedback Guidance Scheduler (FGS) that coordinates these updates across timesteps without additional training. We further extend the framework to multi-object prompts via Delayed Adapter Activation (DAA) and Layer-wise Alternating Guidance (LAG) to prevent premature concept fusion. By overlapping monitoring and denoising through pipelined execution, RADIANCE maintains competitive latency while significantly enhancing the per-sample success rate and effective throughput. Experiments on RareBench and T2I-CompBench demonstrate that RADIANCE consistently enhances compositional alignment and perceptual quality over state-of-the-art baselines.
Zi-Xiang Ni, Bo-Lun Huang, Teng-Fang Hsiao et al.· 0 citations
