GTC ON-DEMAND

 
SEARCH SESSIONS
SEARCH SESSIONS

Search All
 
Refine Results:
 
Year(s)

SOCIAL MEDIA

EMAIL SUBSCRIPTION

 
 

GTC ON-DEMAND

Intelligent Video Analytics
Presentation
Media
Click-based Accelerated Incremental Training of CNNs for Object Detection
Abstract:
In contrast to traditional CNN training with large offline static datasets, some autonomous machine applications will benefit from training in real-time for mid-mission adjustment and correction. This training will occur on live video streams, with a human-in-the-loop. We demonstrate and evaluate a system tailored to performing time-ordered online training (ToOT) in the field, capable of training an object detector on a live video stream with minimal input from a human operator. Online training is conducted entirely on an NVIDIA Jetson TX2 onboard an autonomous machine. We first define training benefit as a metric to measure the effectiveness of a user interaction in a ToOT sequence. We then show that we can obtain annotations for training an object detector from single-point clicks. Furthermore, by exploiting the time-ordered nature of the video stream through object tracking, we can increase the average training benefit of human interactions by several times.
 
Topics:
Intelligent Video Analytics
Type:
Talk
Event:
GTC Silicon Valley
Year:
2018
Session ID:
S8852
Streaming:
Download:
Share: