sample_embedding_folder2/0826165.txt

Ticket Name: TDA2PXEVM: Function to extract the masks from semseg usecase

Query Text:
Part Number: TDA2PXEVM Other Parts Discussed in Thread: TDA2 Hi TI, I have few questions regarding post image processing on tda2. 1. How to use vlib within the tidl_postproclink,c? 2. Which function of vlib can be used to extract the information of individual class mask. What I am looking for is something similar to cv2.inRange function in Vlib? 3. Is there any function to find the contours as well in vlib? Thank you. With best regards, H.M. Owais

Responses:
Hafiz Muhammad Owais said: 1. How to use vlib within the tidl_postproclink,c? I can't find the file tidl_posproclink.c in the SDK. Is there a typo? Also, not sure I understand the question. Assuming that this link is running on DSP, then adding vlib call is like adding any other library call. You need to call into the function you want, and make sure that the header file vlib.h is included and the makefile includes the include path for this. Hafiz Muhammad Owais said: 2. Which function of vlib can be used to extract the information of individual class mask. What I am looking for is something similar to cv2.inRange function in Vlib? There is none in vlib. Hafiz Muhammad Owais said: 3. Is there any function to find the contours as well in vlib? There is none in vlib.

Hi Jesse, Thank you for your reply. I have solved the question 1. I want to be more precise for the rest of questions. Since TI has offered a demo version for semantic segmentation which is semseg. Once we finish the semantic segmentation, let's just say for three classes bicycle, person and background. What I want to do now is to apply image processing further to extract the information regarding whether the person is sitting on the bicycle or not. I believe most of the users need such kind of post image processing for the semantic segmentation. Do you have any solution for the post processing in the vlib? With best regards, Owais

Hi Owals, Unfortunately, no we don’t have such function that would analyze the output of semantic segmentation. However you can train another deep learning network that would analyze the output of the semantic segmentation and detect regions in which there is a person sitting on a bike. Or the original semantic segmentation model could be retrained to detect 4 classes: bicvcle alone, person alone, person sit on bicycle, background. regards, Victor

Hi Victor, I have written post processing algorithm by myself on C which works pretty fine enough :-) without utilizing any vision library. Your suggestion is useful but it would require more time for labeling and training it again. I will keep this suggestion in my mind for other segmentation tasks. Thank you. With best regards, Owais