This paper compares the cross-domain robustness of six established object detector configurations under a unified data protocol, a fixed training budget, and a single evaluation pipeline. All configurations are trained on the union of PASCAL Visual Object Classes (VOC) 2007 and VOC 2012 trainval, comprising 16551 images, under a fixed budget of 18 epochs of source-data exposure, with horizontal flipping at probability 0.5 as the only augmentation and Common Objects in Context (COCO)-pretrained complete detector initialisation. Each configuration is trained under three random seeds. Evaluation covers the VOC 2007 test set, Clipart1k, and Watercolor2k, with all predictions routed through the same pycocotools evaluator. Watercolor degradation is measured against a matched six-category VOC baseline. The Spearman rank correlation between in-domain and Clipart1k average precision at an intersection-over-union threshold of 0.50 is 0.257 across the six configurations. Single Shot Detector (SSD300) ranks fifth in-domain at 0.726 and first on both target domains, whereas Real-Time Detection Transformer (RT-DETR-l) ranks first in-domain at 0.831 and second on both. Observed cross-domain seed variability exceeds in-domain variability for every configuration by factors between 3.4 and 16.2.

Show more

Feed-forward 3D reconstruction makes multi-view geometry accessible, but a video stream often contains more frames than a fixed memory or latency budget can admit. This paper presents a controlled comparison of keyframe sampling for the Visual Geometry Grounded Transformer (VGGT). Five ordinary strategies and one redundancy stress condition are tested on a 116-frame TartanAir trajectory at budgets of 4, 8, and 16 images. Reconstructions are registered to a fixed reference cloud through Sim(3) alignment and evaluated with accuracy, completeness, Chamfer-L1 distance, and F-score. No simple strategy dominates every budget on this trajectory. Sharpness-aware temporal sampling gives the highest observed ordinary F-score at budget 4, but its margin over the random mean lies within the random baseline's standard deviation and is not a deterministic advantage. Appearance diversity gives the highest observed ordinary F-score at budgets 8 and 16. Across three mixed-corruption seeds, its mean F-score falls by 84.7%, whereas sharpness-aware sampling selects no corrupted frames. Threshold and reference-order checks expose further evaluation sensitivity. These conclusions remain limited to one synthetic trajectory and do not establish a new online selector.

Show more

As the crypto-asset market continues to expand, the energy consumption and carbon emissions associated with proof-of-work mechanisms have increasingly entered investors' pricing considerations. Environmental disclosure has therefore become an important source of information for assessing the non-financial risks of digital assets. To examine whether carbon intensity disclosure changes the allocation of investor attention, major crypto assets from 2019 to 2025 are analysed by combining CoinGecko market data, Google Trends search intensity, and publicly available estimates of energy consumption and carbon emissions. Carbon intensity, abnormal search intensity, abnormal trading volume, and realized volatility are constructed, while two-way fixed-effects models and local projections are applied to identify disclosure shocks and their dynamic market consequences. The results indicate that high-carbon assets experience stronger abnormal search attention following disclosure, while the interaction between disclosure events and carbon intensity is significantly positive. A one-standard-deviation attention-misallocation shock increases short-run realized volatility by approximately 0.27 percentage points at its peak, and the abnormal trading response reaches its highest level around the second trading day after disclosure before gradually declining. These patterns indicate that the salience of carbon information first alters investors' information-search structure and subsequently generates short-term trading pressure. More standardized carbon disclosure and the simultaneous presentation of other financial risk indicators may therefore reduce attention distortions associated with highly salient environmental labels and improve pricing efficiency in crypto-asset markets.

Show more

This study evaluated the performance of You Only Look Once (YOLO11n) and Real-Time Detection Transformer (RT-DETR-L) while dealing with car detection tasks based on the Berkeley DeepDrive 100K (BDD100K) dataset and further analyzed their performance under nighttime driving conditions. The models were trained and tested on the homogeneous training and validation sets extracted from BDD100K, with car as the only detection category. Precision, mean Average Precision (mAP), Recall, and inference speed were utilized as the major metrics for evaluation. The results show that RT-DETR-L achieves higher detection precision, recall, and mAP than YOLO11n under both overall and nighttime conditions, while YOLO11n demonstrates a clear advantage in inference speed. Under nighttime conditions, both models exhibit a certain degree of performance degradation compared with the overall validation results, with the mAP@0.5:0.95 metric showing the largest decline for both models. These findings indicate an apparent trade-off between the accuracy and efficiency while executing detection tasks for the two architectures. This study provides a reference for selecting appropriate object detection models under different road and lighting conditions in autonomous driving applications.

Show more