TDA2: How to run SSD based TIDL OD use case in Vision SDK with pre-trained model

Please follow the below steps to Train, Import and run OD use case in VSDK.

1. Download pre-trained model (deploy.prototxt and caffemodel) from gitHub link. And update below parameters in the “deploy.prototxt”. (Otherwise, DSP will run slow)

•     keep_top_k: 20
•     confidence_threshold: 0.15

1. Use the attached import config file with “tidl_model_import.out.exe” available in TIDL  01.01.00.00 release to generate NET and PRM files
2. Build VSDK with below one line change.  (original use case had only 4 classes, but this new model as 21 classes)

• File  : vision_sdk\apps\src\rtos\alg_plugins\objectdetection\objectDrawLink_algPlugin.c
• Line : tempoutPutList->objDesc[tempoutPutList->numObjects].objType = (label - 1)%3;

1. Copy the NET , PRM bin files along with AppImage and Video input files of OD use case. Select the TIDL OD use case to run and see the detection on the Display

 /cfs-file/__key/communityserver-discussions-components-files/791/1307.tidl_5F00_import_5F00_JDetNet_5F00_voc0712.txt

Note :

• The release is mainly intended to give a complete reference for Train, Import, and run of OD use case in VSDK.
• The accuracy of this model will not be as good as the one available in the VSDK3.3 release.  This model is trained using PASCAL VOC dataset.
• The VSDK demo model was trained using TI internal dataset and this is not available publicly so to re-produce the training step, we have trained this reference model using  PASCAL VOC.

Regards,

Kumar.A.D