How to use segnet

duner · October 9, 2020, 4:27pm

How would I use the output of any of the segnet programs? If I just wanted to know, for each pixel, what class was the highest probability, is there any way to do that? The colorful mask is not too easy to use because of all the blending it does, I tried setting alpha to 0 and 255, it did not really help. Is there some other form of output it can generate that I just didnt see?

dusty_nv · October 9, 2020, 5:54pm

Hi @duner, you can get the classID mask by passing in a uint8 image to the segNet.Mask() function.

For example, see this code from the Python segnet example:

github.com

dusty-nv/jetson-inference/blob/022739abc7631fa853ae63550ec091cd838b2105/python/examples/segnet_utils.py#L80


      
          
          
    if self.use_stats:
                  self.class_mask = jetson.utils.cudaAllocMapped(width=self.grid_width, height=self.grid_height, format="gray8")
                  self.class_mask_np = jetson.utils.cudaToNumpy(self.class_mask)
                  
          def ComputeStats(self):
              if not self.use_stats:
                  return
                  
              # get the class mask (each pixel contains the classID for that grid cell)
              self.net.Mask(self.class_mask, self.grid_width, self.grid_height)
          
          
    # compute the number of times each class occurs in the mask
              class_histogram, _ = np.histogram(self.class_mask_np, self.num_classes)
          
          
    print('grid size:   {:d}x{:d}'.format(self.grid_width, self.grid_height))
              print('num classes: {:d}'.format(self.num_classes))
          
          
    print('-----------------------------------------')
              print(' ID  class name        count     %')
              print('-----------------------------------------')

This code gets the class ID mask, and then computes a histogram of all the class ID’s (showing the distribution of classes in the image)

There is also a C++ API for getting the class ID mask, see here: https://github.com/dusty-nv/jetson-inference/blob/022739abc7631fa853ae63550ec091cd838b2105/c/segNet.h#L256

Topic		Replies	Views
Extract info (Class, Location, Etc.) from Segnet.py with custom model and live video feed Jetson Nano jetson-inference	5	346	June 30, 2023
Extract Info(ClassID, Etc) from CityScape Segmentation Jetson TX2 ai-training	21	2255	August 4, 2021
(dusty-nv/jetson-inference) Accessing labels Jetson Nano jetson-inference	4	962	October 18, 2021
Semantic Segmentation_Outdoor Navigation with Segnet and Jetson Nano Jetson TX2 jetson-inference	6	1013	October 18, 2021
Object Name associated with ClassID in net.detect Jetson Nano	3	1773	October 14, 2021
Increased resolution of SegNet class mask to improve data isolation Jetson Nano jetson-inference	3	463	February 2, 2022
Cannot set threshold for segnet Jetson Nano jetson-inference	6	364	February 28, 2023
Inference bug using --stats while running segnet.py when no class is detected Jetson Xavier NX jetson-inference , python , segmentation	7	931	March 9, 2022
Isolating data from images using Segnet Jetson Nano jetson-inference	3	767	December 22, 2021
Extraction in binary image of many body parts using Multi-Human Parsing Jetson Nano jetson-inference , segmentation	2	798	October 15, 2021

How to use segnet

Related topics