Hi team,
Issue:
Run app_tidl_seg from the official demo, replace the segmentation model as your own, image is 512 * 512YUV, results are normal for x86, embedded side results are abnormal (only 1/3 exception below image). The same picture inference on the embedded side is not exactly the same, and is a 1/3 exception below the image.
The same picture, inference on the embedded side, is not exactly the same, and is a 1/3 of the exception below the image.
What's wrong with TI's app_tidl_seg demo, replaced by own models and pictures?
The results are normal for x86, and the image is a valid part of the captured mosaic image, as follows:
Embedded side image, 1/3 of the exception below the image, as follows:
The config file used is as follows, only the sections corresponding to the model and the image have been modified:
# location of conifg
tidl_config ./model/fs_seg/int8/tidl_io_fs_seg_1.bin
# location of network
tidl_network ./model/fs_seg/int8/tidl_net_fs_seg.bin
# location of input files
input_file_path ./input
# location of output files
output_file_path ./out/fs_seg/int8
# start frame Number
start_frame 1
# number of frames
num_frames 2
# input size (width height)
in_size 512 512
# size given to DL network (width height)
# This should should not be less than 4x of input width or height
dl_size 512 512
# size given for display (width height)
# This should should not be less than 4x of input width or height
out_size 512 512
# Maximum number of Semantic Segmentation classes
num_classes 2
# delay in milli seconds (max 2000ms)
delay_in_msecs 0
# Enable or disable output image writing. 1 Enables it , 0 disables it
en_out_img_write 1
# If 1 - Enable display 0 - Disable display
display_option 0
# number of iterations to loop the inputs
num_iterations 1
# interactive input mode 1: yes, 0: no
is_interactive 0
Could you help check this case? Thanks.
Best Regards,
Cherry
