Retrain model depending on the performance of detection model

yaboo-oyabu / CarND-Capstone

MIT License

2 stars 2 forks source link

Retrain model depending on the performance of detection model #24

Closed yaboo-oyabu closed 4 years ago

yaboo-oyabu commented 4 years ago

I retrained the inception v2 coco model.

Dataset: tl_images_for_detection.tgz excluding DriveU (DTLD) dataset
Model: ssd_inception_v2_sdcn_s10000.tgz
Performance: ssd_inception_v2_sdcn_s10000_performance.tgz

All files can be found in https://drive.google.com/drive/u/1/folders/1tzspy7yyseenawGw_viyF2TL87-Tudln

Note that I split train and validation set as 9:1 ratio, and the above performance show the performance on both train and validation set.

@CaoFM @jotredi @antoska20 @xfhg

CaoFM commented 4 years ago

@yaboo-oyabu Looks awesome to me. I couldn't really tell the difference in result between 0.8 vs 0.6. I think because the images from rosbag is free of confusion objects.

yaboo-oyabu commented 4 years ago

@CaoFM @jotredi @antoska20 @xfhg

I re-retrained the model with mixed dataset.

model name: ssd_inception_v2_retrained_2806
performance: ssd_inception_v2_retrained_2806_performance

All files can be found in https://drive.google.com/drive/u/1/folders/1tzspy7yyseenawGw_viyF2TL87-Tudln

More details can be found in https://github.com/yaboo-oyabu/CarND-Capstone/tree/yaboo/real_traffic_lights_collection/models/real_tl_images