Closed yaboo-oyabu closed 4 years ago
I did a basic performance test. It turns out we can't get enough performance without retraining object detection models. Results are located in https://drive.google.com/drive/u/1/folders/1tzspy7yyseenawGw_viyF2TL87-Tudln as tl_light_detection_performance_2020_07_26.tgz.
'ssd_inception_v2_coco_2017_11_17_thres=0.3': pretrained inception v2 model (threshold 0.3) 'ssd_inception_v2_data1_2017_1_17_thres=0.8' retrained model with collected dataset (threshold 0.8) 'ssd_mobilenet_v1_coco_2017_11_17_thres=0.3': pretrained mobilenet v1 model (threshold 0.3) 'ssd_mobilenet_v1_data1-3_2017_11_17_thres=0.8': retrained model with collected + DTLD dataset (threshold 0.8)
I did a basic performance test. It turns out we can't get enough performance without retraining object detection models. Results are located in https://drive.google.com/drive/u/1/folders/1tzspy7yyseenawGw_viyF2TL87-Tudln as tl_light_detection_performance_2020_07_26.tgz.
'ssd_inception_v2_coco_2017_11_17_thres=0.3': pretrained inception v2 model (threshold 0.3) 'ssd_inception_v2_data1_2017_1_17_thres=0.8' retrained model with collected dataset (threshold 0.8) 'ssd_mobilenet_v1_coco_2017_11_17_thres=0.3': pretrained mobilenet v1 model (threshold 0.3) 'ssd_mobilenet_v1_data1-3_2017_11_17_thres=0.8': retrained model with collected + DTLD dataset (threshold 0.8)