Skip to content

Latest commit

 

History

History
143 lines (81 loc) · 3.17 KB

DOCUMENTATION.md

File metadata and controls

143 lines (81 loc) · 3.17 KB

Documentation

Files

These files are for our monocular 3D Tracking pipeline:

requirements.txt installing list of requirements

_init_paths.py import pymot path for tracking evaluation

mono_3d_estimation.py train and test on roipool feature based 3d and tracking

mono_3d_tracking.py compute correlation by IoU and deep feature, and evaluate tracking result via pymot/

motion_lstm.py training script for lstm motion model

scripts/

init.sh builds up the environment

train_gta.sh example training script for gta

train_kitti.sh example training script for kitti

test_gta.sh example testing script for gta

test_kitti.sh example testing script for kitti

model/

dla.py, dla_up.py defines base model of 3D estimation

model.py defines our 3D network architecture

model_cen.py defines our 3D network architecture with an extra node predicts projection of 3D center

motion_model.py defines our lstm architecture

tracker_model.py defines the kalman filter and lstm tracker

tracker_3d.py, tracker_2d.py for computing correlation of objects between frames

loader/

dataset.py data loader for our mono_3d_estimation.py

dla_dataset.py data format for dla.py

gen_dataset.py generate image based features for KITTI and GTA with detection bbox match to gt, BDD format in json files

gen_pred.py prediction placeholder data generation script for BDD format

utils/

config.py defines cfg, including KITTI and GTA

bdd_helper.py helper for generating BDD format

network_utils.py Utility functions for 3D estimation

tracking_utils.py Utility functions for tracking, 3D transformation

plot_utils.py Utility functions for plot 3D bounding boxes and bird's eye view

labels.py defines dataformat for show_labels.py

tools/

convert_estimation_bdd.py converts 3D estimation to BDD format

convert_tracking_bdd.py converts tracking results to BDD format

eval_dep_bdd.py evaluates 3D estimation results using metrics for depth, orientation and center

eval_mot_bdd.py evaluates tracking results using pymot/

plot_tracking.py visualization of 3D tracklets

show_labels.py show ground truth label

visualize_kitti.py visualize and convert format to kitti txt files for devkit/

devkit/ official kitti developement kit to evaluate tracking result

object-ap-eval/ 3D AP evaluation tool

pymot/ multiple object tracking evaluation tool

lib/

make.sh create execution files

Dataset

''' 
Using BDD format with json files
'''

# GTA train
# 447256/457467 frames
gta_data/gta_train_list.json

# For each sequences
gta_data/train/{}_bdd.json

# GTA val with detection available
# 46250/45152 frames
gta_data/gta_val_list.json

# For each sequences
gta_data/val/{}_bdd.json

# GTA test with detection available
# 184459/181034 frames
gta_data/gta_test_list.json

# For each sequences
gta_data/test/{}_bdd.json

Checkpoint

Checkpoint filename is using the following format.

{SESSION}_{DATASET}_checkpoint_{EPOCH}.pth.tar
# For mono_3d_estimation.py
# GTA
./checkpoint/616_gta_checkpoint_030.pth.tar
# KITTI
./checkpoint/623_kitti_checkpoint_100.pth.tar


# For mono_3d_tracking.py
./checkpoint/803_kitti_300_linear.pth