fine-tune object detector

This work proposes a self-supervised feature learning method specifically designed for object detection to address the heavy reliance on large-scale annotated data. By pretraining the feature extractor on unlabeled data and guiding the model to focus on semantically informative object regions, the approach significantly enhances the representational capacity of the detector under limited annotation budgets. Experimental results demonstrate that the proposed method outperforms conventional ImageNet-pretrained models across multiple object detection benchmarks, achieving not only improved detection accuracy but also greater robustness and reliability.

To address domain shift in object detection caused by variations in illumination, viewpoint, and background, this paper presents a systematic review and reconstruction of feature-level domain adaptation methods. We propose a novel conceptual framework—“feature methods as cross-paradigm unifying engines”—and for the first time unify feature-driven mechanisms across adversarial learning, discrepancy minimization, multi-domain collaboration, teacher-student distillation, ensemble modeling, and vision-language alignment, thereby strengthening weakly supervised synthetic-to-real adaptation. Our approach encompasses feature alignment (e.g., MMD, CORAL), reconstruction (e.g., VAE, GAN), transformation (e.g., AdaIN, style transfer), and adversarial gradient reversal. Under a unified evaluation protocol, our method achieves consistent cross-domain detection mAP improvements of 12.3–28.7%, substantially reducing reliance on target-domain annotations and enabling robust deployment in autonomous driving and medical imaging applications.

Existing object detectors often learn task-driven features that rely on shortcut correlations, failing to adequately capture the underlying annotation structure, which limits their generalization, interpretability, and robustness under task shifts or sparse supervision. To address this, this work proposes an annotation-guided feature enhancement framework that explicitly integrates geometric annotation priors into feature learning for the first time. By constructing a dense spatial feature grid and injecting it into the backbone network—where it fuses with the feature pyramid—the method steers region proposal and detection heads toward representations better aligned with annotation structure. Evaluated on wildlife and remote sensing datasets, the approach significantly improves object focus, reduces background sensitivity, and demonstrates superior generalization and data efficiency in weakly supervised and unseen-task settings.

This study investigates the selection of efficient convolutional neural network (CNN) architectures for image classification and object detection under varying task complexity and resource constraints. Through systematic comparisons across five real-world datasets—including binary classification, fine-grained multiclass classification, and object detection tasks—the work evaluates custom CNNs, deep residual networks, and transfer learning models, analyzing the impact of key architectural factors such as network depth and residual connections. The results demonstrate that deeper architectures significantly improve accuracy in fine-grained classification, whereas lightweight pretrained models offer superior efficiency for simpler binary classification tasks. Furthermore, the proposed custom CNN is successfully extended to detect illegally operating tricycles in traffic scenarios, confirming its practical effectiveness in real-world applications.

Current vision-language models (VLMs) struggle to capture fine-grained object details under full-image pretraining, limiting their region-level recognition capability in open-vocabulary object detection. To address this, this work proposes Decoupled Adaptive Training (DAT), which leverages a closed-set detector to generate region-aware pseudo-labels and fine-tunes fewer than 0.8M parameters in the VLM’s visual backbone in a self-supervised manner. This approach enhances local feature alignment while preserving global semantics, requires no inference overhead, and is plug-and-play compatible. By integrating weight interpolation with a collaborative architecture, DAT achieves significant improvements on both seen and novel categories, establishing new state-of-the-art results on COCO and LVIS benchmarks for open-vocabulary object detection.

This study addresses the challenge of precisely identifying high-value images under a limited annotation budget to maximize object detection performance. Moving beyond conventional approaches that rely solely on feature rarity, this work proposes an unlabeled data selection strategy by modeling the relationship between internal model representations and prediction errors. Specifically, it leverages the model's own state and anticipated performance gains to drive active learning. Extensive evaluations demonstrate that the proposed method outperforms rarity-based baselines across most experimental settings and achieves top-ranked performance during long-cycle retraining, substantially improving both annotation efficiency and detection accuracy.

This study systematically compares the performance of three mainstream visual recognition strategies—custom-designed CNNs, fixed pretrained models used as feature extractors, and fine-tuned transfer learning—under real-world conditions. Through controlled experiments across five image classification datasets, the methods are comprehensively evaluated in terms of accuracy, macro F1-score, training time, and parameter count. The work provides the first empirical evidence across diverse real-world domains that transfer learning consistently achieves the best predictive performance, while custom CNNs offer a more favorable trade-off between efficiency and accuracy under computational and memory constraints. These findings offer practical guidance for model selection in applied settings where resource limitations must be balanced against performance requirements.

This work addresses the imbalance in region proposal distributions between base and novel classes in few-shot object detection by proposing a staged proposal refinement strategy. During base-class training, a refinement loss is introduced to enhance the model’s sensitivity to novel classes. In the fine-tuning stage, a lightweight refinement branch is added to the Region Proposal Network (RPN) to generate higher-quality proposals for novel classes, effectively balancing the proposal distribution. Notably, this approach is the first to explicitly tackle distribution shift at the region proposal level. It achieves significant performance gains without increasing inference overhead, surpassing existing methods by 1%–6% across multiple mainstream benchmarks and establishing a new state of the art.

This study addresses the challenge of effectively leveraging large-scale unlabeled images to improve object detection performance under limited annotation budgets. We systematically evaluate three prominent semi-supervised object detection methods—MixPL, Semi-DETR, and Consistent-Teacher—across MS-COCO, Pascal VOC, and a custom Beetle dataset, analyzing their trade-offs among accuracy, model size, and inference latency under varying labeling ratios. For the first time, we reveal consistent patterns of performance degradation as labeled data decreases, both on general-purpose and domain-specific datasets. Our empirical findings provide actionable insights and practical guidance for selecting appropriate semi-supervised approaches in resource-constrained scenarios.

This study addresses the significant performance drop in cross-dataset object detection when transferring models from domain-specific settings—such as driving or aerial imagery—to general-purpose scenarios. It presents the first systematic analysis of such transfers through the lens of “setting specificity,” distinguishing between setting-dependent and setting-agnostic datasets. To disentangle the effects of domain shift and label mismatch, the authors introduce an open-label evaluation protocol that maps predicted categories to target labels via CLIP-based semantic similarity. By integrating both closed- and open-label evaluations, they quantify the contributions of domain-related and label-related errors. Experiments reveal that transfer within similar settings remains stable, whereas cross-setting transfer incurs substantial degradation. The open-label protocol consistently yields modest performance gains, primarily by correcting semantically plausible misclassifications supported by visual evidence.