Pascal VOC과 COCO 이후 주요 객체 검출 벤치마크

다음 표는 Pascal VOC(2007–2012)와 COCO(2014–현재)에 이어 제안된 주요 객체 검출 데이터셋을 비교한 것이다. 각 데이터셋의 출시 연도, 이미지 수, 클래스 수, 어노테이션 종류를 정리하였다.

이들 데이터셋은 Pascal VOC와 MS COCO의 한계를 보완하거나 새로운 난이도·도전 과제를 제시하기 위해 설계되었다.

• LVIS는 long-tail 분포의 1,203개 클래스에 대한 인스턴스 세그멘테이션을 지원하여 희소 클래스 인식 연구를 촉진[1].
• Open Images 시리즈(V4, V6, V7)는 방대한 이미지와 다양한 어노테이션(관계, 멀티모달 내러티브 등)을 제공하여 복합 장면 이해를 목표로 함[2][6].
• Objects365는 365개 클래스, 30M 이상의 바운딩 박스로 대규모 일반 객체 검출을 위한 새로운 기준을 마련[3].
• COCO-ReM은 COCO-2017의 마스크 오류를 교정하여 평가 신뢰도를 높인 리파인된 마스크 벤치마크[4].
• COCO-O는 자연 분포 변화(풍경·스케치·만화 등)에 대한 검출기 강건성을 평가하기 위한 테스트셋[5].

출처
[1] LVIS Dataset – Ultralytics YOLO Docs https://docs.ultralytics.com/datasets/detect/lvis/
[2] The Open Images Dataset V4: Unified image classification, object detection, and visual relationship detection at scale https://arxiv.org/abs/1811.00982v2
[3] Papers with Code – Objects365: A Large-Scale, High-Quality Dataset for Object Detection https://paperswithcode.com/paper/objects365-a-large-scale-high-quality-dataset
[4] [PDF] Benchmarking Object Detectors with COCO: A New Path Forward https://www.ecva.net/papers/eccv_2024/papers_ECCV/papers/06053.pdf
[5] [PDF] A Benchmark for Object Detectors under Natural Distribution Shifts https://openaccess.thecvf.com/content/ICCV2023/papers/Mao_COCO-O_A_Benchmark_for_Object_Detectors_under_Natural_Distribution_Shifts_ICCV_2023_paper.pdf
[6] Open Images V7 https://storage.googleapis.com/openimages/web/factsfigures_v7.html
[7] Benchmarking Object Detectors with COCO: A New Path Forward https://arxiv.org/html/2403.18819v1
[8] Int J Comput Vis (2015) 111:98–136 http://host.robots.ox.ac.uk/pascal/VOC/pubs/everingham15.pdf
[9] COCO dataset https://cocodataset.org
[10] Text and Click inputs for unambiguous open vocabulary instance … https://arxiv.org/html/2311.14822
[11] [Object Detection] 자주 쓰이는 데이터셋, COCO란? – DACON https://dacon.io/en/forum/405930
[12] Table 1. https://pmc.ncbi.nlm.nih.gov/articles/PMC9738404/table/sensors-22-09536-t001/
[13] COCO (Common Objects in Context) Dataset – Papers With Code https://paperswithcode.com/dataset/coco
[14] COCO test-dev Benchmark (Object Detection) – Papers With Code https://paperswithcode.com/sota/object-detection-on-coco
[15] Pascal VOC Dataset: A Technical Deep Dive – viso.ai https://viso.ai/deep-learning/pascal-voc-dataset/
[16] [데이터 셋] COCO dataset – I’m Lim – 티스토리 https://imlim0813.tistory.com/55
[17] ECVA https://www.ecva.net/papers/eccv_2024/papers_ECCV/html/6053_ECCV_2024_paper.php
[18] [PDF] Deep Learning Approaches for Object Detection https://ojs.wiserpub.com/index.php/AIE/article/download/564/398
[19] COCO 2014 Dataset (for YOLOv3) – Kaggle https://www.kaggle.com/datasets/jeffaudi/coco-2014-dataset-for-yolov3
[20] Under review as a conference paper at ICLR 2023 https://openreview.net/pdf/b64642f9ced9d99f97529b07fb5608d3f7f4aedb.pdf
[21] Reconstructing PASCAL VOC https://pdfs.semanticscholar.org/3f0a/6ff8ffbea60709f906b4ff333f951965370a.pdf
[22] [Coding] COCO Dataset 읽기 – velog https://velog.io/@dkdk6638/Pytorch-COCO-Dataset
[23] Under review as a conference paper at ICLR 2021 https://openreview.net/pdf/848af4e0efa1d16b77effbe4ca8d27a18031370f.pdf
[24] Build software better, together https://github.com/topics/pascal-voc
[25] Vehicles-OpenImages Object Detection Dataset https://public.roboflow.com/object-detection/vehicles-openimages
[26] Objects365 Dataset – Ultralytics YOLO Docs https://docs.ultralytics.com/datasets/detect/objects365/
[27] LVIS Dataset – Papers With Code https://paperswithcode.com/dataset/lvis
[28] ultralytics/docs/en/datasets/detect/lvis.md at main – GitHub https://github.com/ultralytics/ultralytics/blob/main/docs/en/datasets/detect/lvis.md
[29] Unified image classification, object detection, and visual relationship … https://arxiv.org/abs/1811.00982
[30] オブジェクト365 https://docs.ultralytics.com/ja/datasets/detect/objects365/
[31] LVIS: A Dataset for Large Vocabulary Instance Segmentation https://deepai.org/publication/lvis-a-dataset-for-large-vocabulary-instance-segmentation
[32] Oggetti365 https://docs.ultralytics.com/it/datasets/detect/objects365/
[33] Efficient, accurate object detection for hundreds of uncommon object classes https://ai.meta.com/blog/efficient-accurate-object-detection-for-hundreds-of-uncommon-object-classes/
[34] Open Images Dataset v5 (Bounding Boxes) – Download https://blog.csdn.net/chengyq116/article/details/103568971
[35] docs/en/datasets/detect/objects365.md · chelseat2023/yolov8 at main https://huggingface.co/spaces/chelseat2023/yolov8/blob/main/docs/en/datasets/detect/objects365.md
[36] LVIS – Dataset Ninja https://datasetninja.com/lvis
[37] Papers with Code – Open Images V4 Dataset https://paperswithcode.com/dataset/open-images-v4
[38] Ultralytics/ultralytics https://dagshub.com/Ultralytics/ultralytics/src/69a2d70a787fc3a60eb32b941d2d6b7a5a0fdfe5/docs/datasets/detect/objects365.md
[39] LVIS Challenge Workshop @ ICCV 2021 https://www.youtube.com/watch?v=uAFHd9MapBM
[40] Open Images V6 https://storage.googleapis.com/openimages/web/factsfigures.html