2D proposals training

[charlesq34]

2D proposals training

[srikanthmalla]

Hi @deli4iled you could use https://github.com/balancap/SSD-Tensorflow for 2D Detections

[skywalker9096]

Hi @charlesq34, In your paper "Frustum PointNets for 3D Object Detection from RGB-D Data", there is a section named "Details on RGB Detector", the base network (half size VGG-16) runs 1.2 ms per image, which seems too fast, because conv1_1 ~ conv5~3 layer of the original VGG16 each cost 0.5ms ~ 1 ms (Pascal P100), did you use other boosting techniques? Or maybe the 1.2 ms per image is divided by batch size?
Furthermore, the Fast-RCNN part runs 30ms per image, too amazing! It's close to SSD speed(~20 ms per image).

[weiliu89]

For the 1st question, the timing is done for a batch of 8 224x224 images, which is ~10ms according to ./build/tools/caffe time, that transfer roughly to 1.2ms per image.

[charlesq34]

(close due to inactivity)