search

team-fusionx / CarND-Capstone

Team FusionX - Capstone Project for Udacity Self-Driving Car Nanodegree
MIT License
8 stars 3 forks source link

New model based only on site images. #141

Closed teeekay closed 6 years ago

teeekay commented 6 years ago

Improve model accuracy by zooming on lights

Tests up to 5 224x224 px clips within the middle strip of the 800x600 px image for tl classification. uses optimized search pattern of clips to reduce number of clips searched on each image based on location of last match.

edufford commented 6 years ago

Other than the small change I requested above, everything worked great during my tests on the simulator and site rosbags so it should be good to go!

teeekay commented 6 years ago

result of set of quick tests run with last Rosbag playback- 1) default single resized image: 0.015 to 0.025 secs GPU time 2) search through clips: generally 0.015 to 0.03 secs GPU time based on 1 to 2 clips analyzed when TL present. between 0.055 to 0.1 secs when no TL present (classify on all 5 clips) 3) batch classify on 5 clips: generally between 0.035 to 0.055 secs GPU time