3D detection and tracking viewer (visualization) for kitti & waymo dataset

Last update: Jan 08, 2023

Overview

3D Detection & Tracking Viewer

This project was developed for view 3D object detection and tracking results. It supports rendering 3D bounding boxes as car models and rendering boxes on images.

Features

Rendering boxes as cars
Captioning box ids(infos) in 3D scene
Projecting 3D box or points on 2D image

Design pattern

This code includes two parts, one for data loading, other one for visualization of 3D detection and tracking results. The overall framework of design is as shown below:

Prepare data

Kitti detection dataset

# For Kitti Detection Dataset         
└── kitti_detection
       ├── testing 
       |      ├──calib
       |      ├──image_2
       |      ├──label_2
       |      └──velodyne      
       └── training
              ├──calib
              ├──image_2
              ├──label_2
              └──velodyne

Kitti tracking dataset

# For Kitti Tracking Dataset         
└── kitti_tracking
       ├── testing 
       |      ├──calib
       |      |    ├──0000.txt
       |      |    ├──....txt
       |      |    └──0028.txt
       |      ├──image_02
       |      |    ├──0000
       |      |    ├──....
       |      |    └──0028
       |      ├──label_02
       |      |    ├──0000.txt
       |      |    ├──....txt
       |      |    └──0028.txt
       |      └──velodyne
       |           ├──0000
       |           ├──....
       |           └──0028      
       └── training # the structure is same as testing set
              ├──calib
              ├──image_02
              ├──label_02
              └──velodyne

Waymo dataset

Please refer to the OpenPCDet for Waymo dataset organization.

Requirements

python3
numpy
vedo
vtk
opencv
matplotlib

Usage

1. Set boxes type & viewer background color

Currently this code supports Kitti (h,w,l,x,y,z,yaw) and Waymo OpenPCDet (x,y,z,l,w,h,yaw) box type. You can set the box type and background color when initializing a viewer as

from viewer.viewer import Viewer

vi = Viewer(box_type="Kitti",bg = (255,255,255))

2. Set objects color map

You can set the objects color map for view tracking results, same as matplotlab.pypot color map. The common used color maps are "rainbow", "viridis","brg","gnuplot","hsv" and etc.

vi.set_ob_color_map('rainbow')

3. Add colorized point clouds to 3D scene

The viewer receive a set of points, it must be a array with shape (N,3). If you want to view the scatter filed, you should to set the 'scatter_filed' with a shape (N,), and set the 'color_map_name' to specify the colors. If the 'scatter_filed' is None, the points will show in color of 'color' arg.

vi.add_points(points[:,0:3],
               radius = 2,
               color = (150,150,150),
               scatter_filed=points[:,2],
               alpha=1,
               del_after_show='True',
               add_to_3D_scene = True,
               add_to_2D_scene = True,
               color_map_name = "viridis")

4. Add boxes or cars to 3D scene

The viewer receive a set of boxes, it must be a array with shape (N,7). You can set the boxes to meshes or lines only, you also can set the line width, conner points. Besides, you can provide a set of IDs(int) to colorize the boxes, and put a set of additional infos to caption the boxes. Note that, the color will set to the color of "color" arg if the ids is None.

vi.add_3D_boxes(boxes=boxes[:,0:7],
                 ids=ids,
                 box_info=infos,
                 color="blue",
                 add_to_3D_scene=True,
                 mesh_alpha = 0.3,
                 show_corner_spheres = True,
                 corner_spheres_alpha = 1,
                 corner_spheres_radius=0.1,
                 show_heading = True,
                 heading_scale = 1,
                 show_lines = True,
                 line_width = 2,
                 line_alpha = 1,
                 show_ids = True,
                 show_box_info=True,
                 del_after_show=True,
                 add_to_2D_scene=True,
                 caption_size=(0.05,0.05)
                 )

You can also render the boxes as cars, the input format is same as boxes.

vi.add_3D_cars(boxes=boxes[:,0:7],
                 ids=ids,
                 box_info=infos,
                 color="blue",
                 mesh_alpha = 1,
                 show_ids = True,
                 show_box_info=True,
                 del_after_show=True,
                 car_model_path="viewer/car.obj",
                 caption_size = (0.1, 0.1)
                )

5. View boxes or points on image

To view the 3D box and points on image, firstly should set the camera intrinsic, extrinsic mat, and put a image. Besides, when adding the boxes and points, the 'add_to_2D_scene' should be set to True.

vi.add_image(image)
vi.set_extrinsic_mat(V2C)
vi.set_intrinsic_mat(P2)

6. Show 2D and 3D results

To show a single frame, you can directly run vi.show_2D(), vi.show_3D(). The visualization window will not close until you press the "Enter" key. Please zoom out the 3D scene by scrolling the middle mouse button backward, and then you can see the point cloud in this window. You can change the viewing angle by dragging the mouse within the visualization window.

To show multiple frames, you can use the for loop, and press the "Enter" key to view a sequence data.

for i in range(len(dataset)):
    V2C, P2, image, boxes = dataset[i]
    vi.add_3D_boxes(boxes)
    vi.add_image(image)
    vi.set_extrinsic_mat(V2C)
    vi.set_intrinsic_mat(P2)
    vi.show_2D()
    vi.show_3D()

3D detection and tracking viewer (visualization) for kitti & waymo dataset

Related tags

Overview

3D Detection & Tracking Viewer

Features

Design pattern

Prepare data

Requirements

Usage

1. Set boxes type & viewer background color

2. Set objects color map

3. Add colorized point clouds to 3D scene

4. Add boxes or cars to 3D scene

5. View boxes or points on image

6. Show 2D and 3D results

Owner

PyTorch implementation of popular datasets and models in remote sensing

The PyTorch re-implement of a 3D CNN Tracker to extract coronary artery centerlines with state-of-the-art (SOTA) performance. (paper: 'Coronary artery centerline extraction in cardiac CT angiography using a CNN-based orientation classiﬁer')

PyTorch implementation of SCAFFOLD (Stochastic Controlled Averaging for Federated Learning, ICML 2020).

A Python toolbox to create adversarial examples that fool neural networks in PyTorch, TensorFlow, and JAX

Source code for the NeurIPS 2021 paper "On the Second-order Convergence Properties of Random Search Methods"

Tooling for GANs in TensorFlow

NaturalProofs: Mathematical Theorem Proving in Natural Language

Reimplementation of Dynamic Multi-scale filters for Semantic Segmentation.

An 16kHz implementation of HiFi-GAN for soft-vc.

Meandering In Networks of Entities to Reach Verisimilar Answers

code for paper"A High-precision Semantic Segmentation Method Combining Adversarial Learning and Attention Mechanism"

The official implementation code of "PlantStereo: A Stereo Matching Benchmark for Plant Surface Dense Reconstruction."

Unsupervised Image Generation with Infinite Generative Adversarial Networks

Add-on for importing and auto setup of character creator 3 character exports.

Codebase for Attentive Neural Hawkes Process (A-NHP) and Attentive Neural Datalog Through Time (A-NDTT)

Volumetric parameterization of the placenta to a flattened template

[CVPR 2022 Oral] Rethinking Minimal Sufficient Representation in Contrastive Learning

ARKitScenes - A Diverse Real-World Dataset for 3D Indoor Scene Understanding Using Mobile RGB-D Data

RuDOLPH: One Hyper-Modal Transformer can be creative as DALL-E and smart as CLIP