AB3DMOT evaluation code

https://github.com/xinshuoweng/AB3DMOT

GitHub - xinshuoweng/AB3DMOT: (IROS 2020, ECCVW 2020) Official Python Implementation for "3D Multi-Object Tracking: A Baseline a

 

Installation

git clone https://github.com/xinshuoweng/AB3DMOT.git

• Python 3.6

cd path/to/AB3DMOT
pip3 install -r requirements.txt

git clone https://github.com/xinshuoweng/Xinshuo_PyToolbox

cd Xinshuo_PyToolbox
pip3 install -r requirements.txt
cd ..

export PYTHONPATH=${PYTHONPATH}:/home/user/workspace/code/AB3DMOT
export PYTHONPATH=${PYTHONPATH}:/home/user/workspace/code/AB3DMOT/Xinshuo_PyToolbox

source ~/.profile
cd path/to/AB3DMOT
source env/bin/activate

 

Run

 def __init__(self, t_sha, gt_path="C:/Users/rosha/OneDrive/Desktop/AVE Lab/AB3DMOT/scripts/KITTI", max_truncation = 0, min_height = 25, max_occlusion = 2, \

        mail=None, cls="car", eval_3diou=True, eval_2diou=False, num_hypo=1, thres=None):

실행 전, evaluation.py 에서 gt_path 부분 수정 필요.

 

$python scripts\KITTI\evaluate.py pointrcnn_Car_val_H1 1 3D 0.25

 

=========evaluation with confidence threshold 7.286145, recall 0.025000=========
 sMOTA   MOTA   MOTP    MT     ML     IDS  FRAG    F1   Prec  Recall  FAR     TP    FP    FN
0.9266 0.0232 0.8058 0.0000 0.9444     0     0 0.0501 0.9756 0.0257 0.0030    80     2  3034
================================================================================
=========evaluation with confidence threshold 2.659125, recall 0.050000=========
 sMOTA   MOTA   MOTP    MT     ML     IDS  FRAG    F1   Prec  Recall  FAR     TP    FP    FN
0.7593 0.0380 0.7599 0.0333 0.8889     0     0 0.0851 0.9329 0.0446 0.0152   139    10  2980
================================================================================
=================evaluation: best results with single threshold=================
Multiple Object Tracking Accuracy (MOTA)                                  0.0380
Multiple Object Tracking Precision (MOTP)                                 0.7599
Multiple Object Tracking Accuracy (MOTAL)                                 0.0380
Multiple Object Detection Accuracy (MODA)                                 0.0380
Multiple Object Detection Precision (MODP)                                0.9922

Recall                                                                    0.0446
Precision                                                                 0.9329
F1                                                                        0.0851
False Alarm Rate                                                          0.0152

Mostly Tracked                                                            0.0333
Partly Tracked                                                            0.0778
Mostly Lost                                                               0.8889

True Positives                                                               139
Ignored True Positives                                                        11
False Positives                                                               10
False Negatives                                                             2980
Ignored False Negatives                                                      538
ID-switches                                                                    0
Fragmentations                                                                 0

Ground Truth Objects (Total)                                                3657
Ignored Ground Truth Objects                                                 549
Ground Truth Trajectories                                                     96

Tracker Objects (Total)                                                      150
Ignored Tracker Objects                                                        1
Tracker Trajectories                                                          28
================================================================================
========================evaluation: average over recall=========================
 sAMOTA  AMOTA  AMOTP
0.0421 0.0015 0.0391
================================================================================

 

Input and Output

Data type

KITTI dataset(17dim)

frame  tracklet_id  objectType  truncation  occlusion  alpha  x1  y1  x2  y2  h  w  l  X  Y  Z  ry

#Values    Name      Description
----------------------------------------------------------------------------
   1    frame        Frame within the sequence where the object appearers
   1    track id     Unique tracking id of this object within this sequence
   1    type         Describes the type of object: 'Car', 'Van', 'Truck',
                     'Pedestrian', 'Person_sitting', 'Cyclist', 'Tram',
                     'Misc' or 'DontCare'
   1    truncated    Integer (0,1,2) indicating the level of truncation.
                     Note that this is in contrast to the object detection
                     benchmark where truncation is a float in [0,1].
   1    occluded     Integer (0,1,2,3) indicating occlusion state:
                     0 = fully visible, 1 = partly occluded
                     2 = largely occluded, 3 = unknown
   1    alpha        Observation angle of object, ranging [-pi..pi]
   4    bbox         2D bounding box of object in the image (0-based index):
                     contains left, top, right, bottom pixel coordinates 
   3    dimensions   3D object dimensions: height, width, length (in meters)
   3    location     3D object location x,y,z in camera coordinates (in meters)
   1    rotation_y   Rotation ry around Y-axis in camera coordinates [-pi..pi]

https://github.com/JonathonLuiten/TrackEval/blob/master/docs/KITTI-format.txt

 

input

/results/KITTI\pointrcnn_Car_val_H1\data_0\0001.txt

/results/KITTI\pointrcnn_Car_val_H1\data_0\0016.txt

output

C:/Users/rosha/OneDrive/Desktop/AVE Lab/AB3DMOT/scripts/KITTI\label\0001.txt

C:/Users/rosha/OneDrive/Desktop/AVE Lab/AB3DMOT/scripts/KITTI\label\0016.txt