search

mathmanu / caffe-jacinto-models

This repository has moved. The new link can be obtained from https://github.com/TexasInstruments/jacinto-ai-devkit
63 stars 47 forks source link

any docs about SSD Object detection ? #1

Open ianz27 opened 6 years ago

ianz27 commented 6 years ago

is there any information about ssd ? like trained model and performance

mathmanu commented 6 years ago

Some clarifications:

  1. Tried out some of my custom configurations - and it trains well and produces good results. However I have not tried to reproduce the original mAP of the VGG based SSD model published by Wei Liu. It should be possible as there are no changes in this port that alter the functionality.

  2. Have not tried with FP16 data type. I tried only with the default (float) data types.

  3. NVIDIA/caffe allows forward, backward and math types to be different. However BVLC/caffe doesn't support this. This porting over may not be perfect - so, these SSD layers may not work correctly if you use different types (forward, backward, math) in these SSD layers. Fixing this need careful attention to be spend on the Backward functions of some of these SSD layers.

ianz27 commented 6 years ago

Hi @mathmanu I think it's hard to reproduce the original mAP because the network(jdetnet21v2) is so small. I only get 69% mAP(initial), and the original mAP is 79.8%. Could you share your configurations and results ?

mathmanu commented 6 years ago

jdetnet21v2 is small because its targeted for low power embedded devices.

What I meant is that if you train with the original VGG based configuration published at https://github.com/weiliu89/caffe/tree/ssd, you should be able to re-produce the original results. If you have some time, could you try that?

sagartesla commented 6 years ago

where can i get architecture for jdetnet21v2 ? deploy.prototxt and etc

mathmanu commented 6 years ago

After cloning this repository, in your bash prompt, change directory to the caffe-jacinto-models/scripts folder, run the following command: ./train_image_object_detection.sh

The prototxt files will be generated in the folder caffe-jacinto-models/scripts/training

For any other details of training, please consult the documentation given in in the original repository: https://github.com/weiliu89/caffe/tree/ssd The same applies when you train using the fork tidsp/caffe-jacinto as well.