Visual Object Tracking Benchmarks

VastTrack

VastTrack: Vast Category Visual Object Tracking. NeurIPS'24

• The number of sequences: 50610
• The average sequence length: 83 frames
• The number of object categories: 2115
• 특징: 짧지만 많은 수의 video sequences와 다양한 object categories

VOTS2024

The Second Visual Object Tracking Segmentation VOTS2024 Challenge Results. ECCVW'24

• The number of sequences: 144
• The average sequence length: 2000 frames (min: 63, max: 10700, median: 1810)
• The number of targets: 341
• The average number of targets: 2.37 (min: 1, max: 8, median: 2)
• 특징
  • VOT2022와 달리 아예 새롭게 제작
  • Short-term, long-term, single and multi-target tracking scenarios를 커버
  • Target이 fov를 나갔다가 reappear하는 경우도 있음 (93 sequences, 168 targets). 여러번 reappear 함 (median: 3, max: 23, median absence length: 18)

VOT2022

The Tenth Visual Object Tracking VOT2022 Challenge Results. ECCVW'22

• The number of sequences: 62 (VOT2020에서 간단한건 제외되고 challenging한 것 추가)

VOT2020

The Eighth Visual Object Tracking VOT2020 Challenge Results. ECCVW'20

• The number of sequences: 60
• Metric: Expected Average Overlap (EAO)

LaSOT

LaSOT: A High-Quality Benchmark for Large-Scale Single Object Tracking. CVPR'19

• The number of sequences: 1,400
• The number of object categories: 70
• The average sequence length: 2,500 frames
• Train/Test: 1,120/280 sequences (16/4 sequences for each category)

LaSOT_ext

LaSOT: A High-Quality Large-Scale Single Object Tracking Benchmark. IJCV'21

• An extension to the original LaSOT dataset
• The number of additional sequences: 150
• The number of new object categories: 15
• These new sequences are specifically designed to focus on occlusions and variations in small objects, which is more challenging.
• The standard protocol is to evaluate the models trained on LaSOT directly on the LaSOT_ext.

GOT-10k

GOT-10k: A Large High-Diversity Benchmark for Generic Object Tracking in the Wild. PAMI'19

• The number of sequences: 10,000
• The number of object categories: 560
• The number of motion patterns: 80+
• A key aspect of GOT-10k is its one-shot evaluation protocol, which requires trackers to be trained exclusively on the designated training split, with 170 sequences reserved for testing.

TrackingNet

Trackingnet: A Large-Scale Dataset and Benchmark for Object Tracking in the Wild. ECCV'18

• The number of sequences: 30,643
• The average sequence length: 471 frames
• Training/Test: 30,132/511 sequences

NFS

Need for Speed: A Benchmark for Higher Frame Rate Object Tracking. ICCV'17

• The number of sequences: 100
• The number of frames: 380,000 (240 fps)
  • VOT works에서는 주로 30 fps로 downscale해서 사용하는듯 함 [SAMURAI].

OTB100

Object Tracking Benchmark. PAMI'15

• The number of sequences: 100
• The average sequence length: 590 frames