SHI-Yu-Zhe / Benchmarks-for-Single-Object-Visual-Tracking

Exploring through 7 popular datasets for visual object tracking, including OTB, UAV, VOT, LaSOT, NFS, TrackingNet and GOT-10k.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Benchmarks: Single-Object Visual Tracking

Better Viewed at local markdown readers like Typora.

Dataset Overview

Dataset Source Domain
OTB100 Exist Datasets(e.g. FERET) People
VOT2018 Exist Datasets(e.g. OTB) People
UAV123 (UAV+HD Camera; UAV+miniC; Simulator) Vehicles
TrackingNet Youtube(YT-BB) Diverse
NFS2017 Embedded Vision/Mobile High fps
LaSOT Youtube (70 Categories defined by ImageNet) Diverse
GOT-10k Youtube (retrived by semantic information) Diverse
Dataset Novelty
OTB100 The basic Performance evaluation (applied in other datasets)
VOT2019 Performance evaluation; reinitialization at tracking failures;Bounding box rotation
UAV123 Aerial video dataset; UE4-based Simulator
TrackingNet Dataset generate dense annotations from coarse ones
NFS2017 Dataset obtained from consumer devices, in high fps
LaSOT Large-scale dataset, Natural Language Description of every sequence
GOT-10k Tremendous Diverse Classes; Annotation of Visible Ratios; Challenge Indicators
Dataset Videos Classes Min frames Mean frames Median frames Max frames Total frames frame rate
OTB100 100 22 71 590 393 3872 59K 30fps
VOT2019 60 30 41 356 293 1500 21K 30fps
UAV123 123 9 109 915 882 3085 113K 30fps
TrackingNet 30643 21 - 470 - - 14.42M 30fps
NFS2017 100 33 169 3830 2448 20665 383K 240fps
LaSOT 1400 70 1000 2506 2053 11397 3.52M 30fps
GOT-10k 10000 563 4 150 - 1480 56K 10fps

Evaluation Methodology: Precision, Accuracy, Robustness

P -- Precision

Calculating center position error by euclidean distance between result center and annotation center.

Visible Ratio (GOT-10k)

Visible Ratio is a continuous labeling for object occlusion. A visible ratio is a percentage indicating the approximate proportion of an object that is visible.

IoU->AOS+AUC: Intersection over Union -- Accuracy

$$ S=\frac{|r_t\cap r_0|}{|r_t\cup r_0|} $$

S denotes Average Overlap Score.

Corresponding to the sampled overlap threshold, Area Under Curve can be generated by average.

OPE+TRE+SRE -- Robustness

The basic factor to evaluate robustness of a tracker is the number of times the tracker failed.

One Pass Evaluation: Test each tracker on each sequence once.

Spatial Robustness Evaluation: Adding 8 spatial shifts(4 center shifts, 4 corner shifts) and 4 scale variations to bounding boxes in 10% frames of entire sequence. Test each tracker on each sequence for 12 times.

Temporal Robustness Evaluation: Tracker running starts from different frames. Test each tracker on each sequence for 20 times.

OPER+SRER -- Accuracy+Stability+Robustness

One Pass Evaluation with Restart: use ground truth to restart a tracker after its failure.

Spatial Robustness Evaluation with Restart: add spatial perturbation(noise).

Reinitialization (VOT)

The idea is developed from OPE. If the tracker loses tracking for continuous 5 frames, it is restarted with ground truth. To reduce the bias after ground truth, the first 10 frames ,i.e. Burn in, is excluded from performance evaluation.

Average per Frame (VOT): per-frame Accuracy+Robustness

$$ P_A=\frac{1}{N}\sum_1^Naccuracy(tracker_i,frame_t) $$

$$ P_R=\frac{1}{N}\sum_1^N(robustness(tracker_i,frame_t)) $$

Dataset Evaluation Methodology
OTB100 P;IoU->AOS+AUC; OPE+OPER; TRE+SRE+SRER; Vitual runs
VOT2019 P;IoU->AOS+Re-initialized(N=5)+(Burnin=10); Per-frame A/R
UAV123 P;IoU->AOS+AUC; OPE+OPER; SRE; Simulator
TrackingNet P; IoU->AOS+AUC; OPE with normalization
NFS2017 IoU->AOS+AUC; Improved-Accuracy/Low-fps-Accuracy(240->30)
LaSOT P; IoU->AOS+AUC; OPE
GOT-10k P; Visible Ratio; IoU->AOS+AUC; OPE

Challenges

Challenging tasks in tracking:

Camera Motion: Abrupt motion of the camera; (CM)

In-Plane Rotation: The target rotates in the image plane; (IPR)

Out-of-Plane Rotation: The target rotates out of the image plane; (OPR)

Deformation: Non-rigid object deformation; (DEF)

Fully Occluded: The target is fully occluded; (FOC)

Partially Occluded: The target is partially occluded; (POC)

Illumination Variation: The illumination of the target changes significantly; (IV)

Out of View: Some portion of the target leaves the camera field of view; (OV)

Viewpoint Change: Viewpoint affects target appearance significantly. (VC)

Scale Variation: The ratio of bounding box area is outside the range [0.5, 2] after 1s; (SV)

Motion Blur: The target region is blurred due to the motion of target or camera; (MB)

Aspect Ratio Change: The ratio of bounding box aspect ration is outside the range [0.5, 2] after 1s; (ARC)

Low Resolution: At least one ground truth bounding box has less than 1000 pixels; (LR)

Fast Motion: The motion of the bounding box is larger than the size of the bounding box; (FM)

Background Clutter: The background near the target has similar appearance as the target; (BC)

Similar Object: There are objects of similar shape or same type near the target. (SO)

Challenge Indicators (GOT-10k)

Occlusion/Truncation: $$ OcclusionIndicator=1-v $$ where v denotes for labeled visible ratio.

Scale Variation: $$ ScaleVariationIndicator=\max{s_i/s_{i-T},s_{i-T}/s_i} $$ where $$ s_i=\sqrt{w_i\times h_i} $$ Aspect Ratio Change: $$ AspectRatioChangeIndicator=\max{s_i/s_{i-T},s_{i-T}/s_i} $$ where $$ s_i=h_i/w_i $$ Fast Motion: $$ d_i=\frac{|p_i-p_{i-1}|2}{\sqrt{s_i\times s{i-1}}} $$ where p denotes the object center location and $$ s_i=\sqrt{w_i\times h_i} $$ Illumination Variation: $$ u_i=|c_i-c_{i-1}|_1 $$ where c is average object color, RGB value normalized to [0, 1].

Low Resolution: $$ LowResolutionIndicator=s_i/s^{median}, s_i\leq s^{median} $$ where $$ s_i=\sqrt{w_i\times h_i} $$

OTB100 VOT2019 UAV123 TrackingNet NFS2017 LaSOT GOT-10k
CM 1 1 1 1
IPR 1 1 1
OPR 1 1
DEF 1 1 1 1 1
FOC 1 1 1 1 1 1 1
POC 1 1 1 1 1 1 1
IV 1 1 1 1 1 1 1
OV 1 1 1 1 1
VC 1 1 1
SV 1 1 1 1 1 1 1
MB 1 1 1 1
ARC 1 1 1 1 1
LR 1 1 1 1 1 1
FM 1 1 1 1 1 1
BC 1 1 1 1 1 1
SOB 1 1

About

Exploring through 7 popular datasets for visual object tracking, including OTB, UAV, VOT, LaSOT, NFS, TrackingNet and GOT-10k.