DriveWorks SDK Reference
3.0.4260 Release
For Test and Development only

perception/object/camera/drivenet/docs/drivenet_mainsection.md
Go to the documentation of this file.
1 # Copyright (c) 2019-2020 NVIDIA CORPORATION. All rights reserved.
2 
3 @page drivenet_mainsection DriveNet
4 
5 @note SW Release Applicability: This module is available in **NVIDIA DRIVE Software** releases.
6 
7 ## About This Module
8 
9 DriveNet is the NVIDIA proprietary object detection deep neural network: **DriveNet**.<br>
10 DriveWorks comes with four DriveNet models:
11 - front camera model: ::DW_DRIVENET_MODEL_FRONT
12 - side camera mode: ::DW_DRIVENET_MODEL_SIDE
13 - stateful temporal model with urgency prediction: ::DW_STATEFUL_TEMPORAL_DRIVENET_MODEL
14 - stateless temporal model with urgency prediction: ::DW_STATELESS_TEMPORAL_DRIVENET_MODEL
15 
16 ### Inputs
17 
18 DriveNet consumes FP16 planar frames with a resolution of 480x240 (side camera model) or 960x480 (front camera model) pixels from AR0231 cameras (revision >= 4).
19 
20 @note resizing the input frame is internally handled by @ref object_description1 as part of the normal workflow.
21 
22 ::DW_DRIVENET_MODEL_FRONT model is trained to support any of the following six camera configurations:
23 - Front camera location with a 30&deg; field of view
24 - Front camera location with a 60&deg; field of view
25 - Front camera location with a 120&deg; field of view
26 - Rear camera location with a 60&deg; field of view
27 - Front-left camera location with a 120&deg; field of view
28 - Front-right camera location with a 120&deg; field of view
29 - Rear-left camera location with a 120&deg; field of view
30 - Rear-right camera location with a 120&deg; field of view
31 
32 @warning Current models are not fully trained on the 30&deg; field of view camera, resulting in functions being less validated.
33 
34 ::DW_DRIVENET_MODEL_SIDE is trained to support side camera location with a 120&deg; field of view
35 
36 Temporal models (::DW_STATEFUL_TEMPORAL_DRIVENET_MODEL and ::DW_STATELESS_TEMPORAL_DRIVENET_MODEL
37 ) support front camera with 60&deg; field of view.
38 
39 ### Outputs
40 
41 DriveNet allows to detect a range of different objects as 2D bounding boxes.<br>
42 The following classes are currently supported on all models:
43 - Car
44 - Bicycle
45 - Pedestrian
46 - Traffic sign
47 - Traffic light
48 
49 @note Trucks are also detected, however they are currently classified as cars.
50 @note Motorbikes are classified as bicycles.
51 
52 ![Multiclass object detector implemented using DriveNet](sample_drivenet.png)
53 
54 Detections coming from DriveNet are associated with:
55 - a confidence value in the [0,1] range<br>
56 - a depth value [m]
57 - an urgency value [1/s] which is the inverse of time to collision<br>
58  Frame rate is internally considered for the computation of the urgency.
59 
60 @note `dwDriveNet_getClassProperties()` can be used to query if depth and urgency are available a particular model / class.
61 
62 ### Additional Information
63 
64 While by default inference on DriveNet models is executed on GPUs, it is also possible to configure them to use<br>
65  the Deep Learning Accelerators available on Drive AGX boards.
66 
67 @warning Currently, the DriveNet DNN has limitations that could affect its performance:
68 - It is optimized for daytime, clear-weather data. As a result, it
69  does not perform well in dark or rainy conditions.
70 - It is trained primarily on data collected in the United States.
71  As a result, it may have reduced accuracy in other locales,<br>
72  particularly for road sign shapes that do not exist in the U.S.
73 
74 @note For more information on running inference using DriveNet see @ref object_mainsection.
75 
76 ## Relevant Tutorials
77 
78 - @ref drivenet_usecase1
79 - @ref object_usecase1
80 
81 ## APIs
82 
83 - @ref drivenet_group