search

NVIDIA-AI-IOT / redtail

Perception and AI components for autonomous mobile robotics.
BSD 3-Clause "New" or "Revised" License
1.01k stars 344 forks source link

Output seems to be not right #100

Open mohanen opened 6 years ago

mohanen commented 6 years ago

Hi

I ran the stereoDNN in TX2 with the image given in the sample app and the output doesn't seem to be right.

nvidia@tegra-ubuntu:~/mohanen/stereoDNN/bin$ ./nvstereo_sample_app nvsmall 513 161 ../models/NVTiny/TensorRT/trt_weights_fp16.bin ../sample_app/data/img_left.png ../sample_app/data/img_right.png disp fp16

img_left img_right

Output disp

Can anyone let me know what seems to be the problem here ..

Thanks and Regards, Mohanen B

Alexey-Kamenev commented 6 years ago

At the moment, the only network that correctly handles FP16 is ResNet18_2D.

The history is a little bit complicated: Stereo DNN code was originally released for TensorRT 3.0 where all networks were working fine, though not fully utilizing FP16 potential. In TRT 4.0, there were some changes which enabled proper FP16 support (and close to 2x speedup) in plugins given the plugins opt-in. We did not have time to update all of our plugins, so other networks will not work properly in FP16.

mohanen commented 6 years ago

Ok, I am able to get the output in fp32 with NVTiny

XKXV-rose commented 5 years ago

hi, i have a problem. when i execute nvstereo_sample_app, there is a error: cannot execute binary file. i have do "chmod 777 file" . i cannot solve it, could help me, thank u