Small object detection on highways via balance feature fusion and task-specific encoding network

Abstract

Detecting small objects on highways is a novel research topic. Due to the small pixel of objects on highways, traditional detectors have difficulty in capturing discriminative features. Additionally, the imbalance of feature fusion methods and the inconsistency between classification and regression tasks lead to poor detection performance on highways. In this paper, we propose a balance feature fusion and task-specific encoding network to address these issues. Specifically, we design a balance feature pyramid network (FPN) to integrate the importance of each layer of feature maps and construct long-range dependencies among them, thereby making the features more discriminative. In addition, we present task-specific decoupled head, which utilizes task-specific encoding to moderate the imbalance between the classification and regression tasks. As demonstrated by extensive experiments and visualizations, our method obtains outstanding detection performance on small object detection on highways (HSOD) dataset and AI-TOD dataset.

References

[1] WangC Y, BochkovskiyA, LiaoH M. Yolov7: trainable bag-of-freebies sets new state-of-the-art for real-time object detectors[C], 2023, New York, IEEE: 7464-7475

[2] CarionN, MassaF, SynnaeveG, et al.. End-to-end object detection with transformers[C], 2020, Berlin, Heidelberg, Springer: 213-229

[3] ChanS X, LiuP, ZhangZ. Webox: locating small objects from weak edges[J]. Optoelectronics letters, 2021, 17(6):349-353

[4] KimS, KookH, SunJ, et al.. Parallel feature pyramid network for object detection[C], 2018, Berlin, Heidelberg, Springer: 239-256

[5] LiuS, QiL, QinH, et al.. Path aggregation network for instance segmentation[C], 2018, New York, IEEE: 8759-8768

[6] WuY, ChenY P, YuanL, et al.. Rethinking classification and localization for object detection[C], 2020, New York, IEEE: 10183-10192

[7] GE Z, LIU S T, WANG F, et al. YOLOX: exceeding YOLO series in 2021[EB/OL]. (2021-07-18) [2023-06-24]. https://arxiv.org/abs/2107.08430.

[8] BOCHKOVSKIY A, WANG C, LIAO H M. Yolov4: optimal speed and accuracy of object detection[J]. (2020-04-23) [2023-06-24]. https://arxiv.org/abs/2004.10934.

[9] ZhangY F, RenW Q, ZhangZ, et al.. Focal and efficient IOU loss for accurate bounding box regression[J]. Neurocomputing, 2022, 506: 146-157

[10] WangJ W, YangW, GuoH, et al.. Tiny object detection in aerial images[C], 2021, New York, IEEE: 3791-3798

[11] LinT, MaireM, BelogieS, et al.. Microsoft COCO: common objects in context[C], 2014, Berlin, Heidelberg, Springer: 740-755

[12] GirshickR. Fast R-CNN[C], 2015, New York, IEEE: 1440-1448

[13] CaiZ W, VasconcelosN. Cascade R-CNN: delving into high quality object detection[C], 2018, New York, IEEE: 6154-6162

[14] ZHOU X Y, WANG D Q, KR Ä HENB Ü HI P. Objects as points[EB/OL]. (2019-04-25) [2023-06-24]. https://arxiv.org/abs/1904.07850v1.

[15] LuX, LiB Y, YueY X, et al.. Grid R-CNN[C], 2019, New York, IEEE: 7363-7372

[16] ZhangS F, ChiC, YaoY Q, et al.. Bridging the gap between anchor-based and anchor-free detection via adaptive training sample selection[C], 2020, New York, IEEE: 9759-9768

[17] DaiX Y, ChenY P, XiaoB, et al.. Dynamic head: unifying object detection heads with attentions[C], 2021, New York, IEEE: 7373-7382virtual

[18] FengC J, ZhongY J, GaoY, et al.. TOOD: task-aligned one-stage object detection[C], 2021, New York, IEEE: 3490-3499

[19] ZhangH, LiF, LiuS L, et al.. DINO: DETR with improved denoising anchor boxes for end-to-end object detect ion[C], 2023, New York, IEEE

[20] XuC, WangJ W, YangW, et al.. Dot distance for tiny object detection in aerial images[C], 2021, New York, IEEE: 1192-1201virtual

[21] WANG J W, XU C, YANG W, et al. A normalized Gaussian Wasserstein distance for tiny object detection[EB/OL]. (2021-10-26) [2023-06-24]. https://arxiv.org/abs/2110.13389.

[22] REDMON J, FARHADI A. Yolov3: an incremental improvement[EB/OL]. (2018-04-08) [2023-06-24]. https://arxiv.org/abs/1804.02767.

[23] KimK, LeeH S. Probabilistic anchor assignment with iou prediction for object detection[C], 2020, Berlin, Heidelberg, Springer: 355-371

[24] QiaoS Y, ChenL, YuilleA. DetectoRS: detecting objects with recursive feature pyramid and switchable atrous convolution[C], 2021, New York, IEEE: 10213-10224virtual