@Article{cmc.2026.087886,
AUTHOR = {Seongwook Lee, Seungyeong Kim, Jaemin Oh, L. Minh Dang, Hyeonjoon Moon},
TITLE = {ReMiss: Missed-Object History for Data Scheduling in Driving-Scene Object Detection},
JOURNAL = {Computers, Materials \& Continua},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/cmc/online/detail/28371},
ISSN = {1546-2226},
ABSTRACT = {Average precision (AP) is an aggregate detector metric and does not identify which annotated ground-truth instances remain unmatched across repeated evaluations during training. Existing hard example mining and sampling strategies usually define difficulty at the image, proposal, class, or loss level and do not retain recurrent object-level failures after post-processing as a data-scheduling signal. ReMiss uses missed-object history as a training-time data scheduling signal without changing detector architecture or loss. MissBank records each ground-truth instance’s miss status, miss-count history, and localization and confidence diagnostics. Hard Replay (HR) assigns priorities to full images containing currently missed objects based on miss counts and matching gaps, then schedules those images into later epochs under a replay budget. We evaluate ReMiss on KITTI, BDD100K, Cityscapes, and nuImages with fully convolutional one-stage object detection (FCOS), Faster R-CNN, and detection transformer with improved denoising anchor boxes (DINO), yielding 12 dataset–detector pairs. Across the full replay-ratio sweep, all 36 dataset–detector–ratio configurations improve upon the No-HR diagnostic baseline, with a minimum gain of <math id="mml-ieqn-1"><mo>+</mo><mn>0.22</mn></math> AP (sign test, <math id="mml-ieqn-2"><mi>p</mi><mo>&lt;</mo><msup><mn>10</mn><mrow><mo>−</mo><mn>10</mn></mrow></msup></math>); under a replay ratio fixed in advance at <math id="mml-ieqn-3"><mn>5.0</mn><mi mathvariant="normal">%</mi></math>, all 12 pairs improve by <math id="mml-ieqn-4"><mo>+</mo><mn>0.64</mn></math>–<math id="mml-ieqn-5"><mo>+</mo><mn>2.17</mn></math> AP. Selecting the HR setting by validation mean average precision (mAP), reported at <math id="mml-ieqn-6"><msub><mtext>mAP</mtext><mrow><mn>50</mn><mo>:</mo><mn>95</mn></mrow></msub></math>, yields <math id="mml-ieqn-7"><mo>+</mo><mn>1.03</mn></math>–<math id="mml-ieqn-8"><mo>+</mo><mn>2.23</mn></math> AP. These results support missed-object history as an architecture- and loss-agnostic scheduling signal across the evaluated detector families, with a measured 13%–27% increase in training time over the training-only baseline and no additional inference cost.},
DOI = {10.32604/cmc.2026.087886}
}