Short-horizon collision–risk warning can use instantaneous assessment, forecast-then-assess, or direct prediction from encounter histories. We evaluate the third route with the Anticipatory TRAffic-context Collision–risk Transformer (ATRACT), which estimates the maximum near-future fuzzy Collision Risk Index (CRI) directly from own-ship kinematic and collision-geometry histories without trajectory rollout. Experiments use ten days of Automatic Identification System (AIS) data from the Danish straits, comprising 1.63 million decision windows. Using a vessel-day-track partition and five random initializations, ATRACT attains an area under the receiver operating characteristic curve (ROC-AUC) of , compared with for the evaluated forecast-then-CRI baseline (VCRF) and 0.729 for instantaneous CRI. Thresholds fixed at a nominal 5% false-alarm rate (FAR) on disjoint calibration tracks yield test FARs of 4.98% and 5.38% for ATRACT and VCRF. Although seed-0 track-bootstrap intervals include zero at every evaluated offset, ATRACT shows 4.4–10.8 percentage points higher mean pre-onset detection across 0.5–4 min over five random initializations; VCRF detects more individual high-risk windows at this operating point. Four direct-risk encoders achieve a narrow ROC-AUC range of 0.813–0.819. These results support direct temporal interaction modeling as a viable route while limiting conclusions to the evaluated forecast-first implementations.
Image compression has evolved from human-centric perceptual coding toward support for diverse machine vision applications, requiring modern codecs to serve both human viewing and downstream tasks in closed-set settings (where target tasks are incorporated during training) and open-set settings (where previously unseen tasks arise at test time). While recent learned compression methods jointly optimize perceptual quality and closed-set task performance, they often fail to generalize to unseen open-set tasks due to fixed training assumptions and objectives. Our prior work, LoRA-comp (Low-Rank Adaptation Compression), effectively addresses open-set challenges via instance-specific test-time fine-tuning (TTFT) without requiring task-specific pre-training. Nevertheless, its fixed LoRA architecture, which assigns a uniform rank across all layers, often leads to suboptimal instance-level performance. Moreover, allocating the same number of training epochs to every instance introduces unnecessary encoding-time overhead. To address these challenges, we propose Instance-Specific Early Stopping with Dynamic Rank Adaptation (ISER), which extends LoRA-comp. Building upon the LoRA-comp–based instance-specific adaptation framework, ISER introduces (i) instance-specific early stopping (ISES) combined with a multi-scale training strategy (MSTS) to reduce TTFT overhead and (ii) instance-specific dynamic rank adaptation (ISRA) to tailor the LoRA architecture per instance. Experiments demonstrate that ISER consistently outperforms competing methods. Compared to LoRA-comp, ISER achieves up to a 7% BD-Rate improvement and up to a 44% reduction in encoding time. Moreover, ISER achieves up to a 98% relative improvement in BD-Rate gain and up to a 19.2% reduction in decoding time over TransTIC.