How to incorporate a very wide video input (1920x720) into deepstream inference?

How to infer very wide video (1920x720) through resnet10.

• GPU - 3050Ti Mobile
• DeepStream Version -6.1
• TensorRT Version - 8.4.2-1
• NVIDIA GPU Driver Version - 515.48.07
• Issue Type - question
• How to reproduce the issue ?
I am using deepstream/samples/configs/deepstream-app/source1_usb_dec_infer_resnet_int8.txt. Just have a source of wide video instead of usb camera. Inference is working but as the video is wide, inference is not accurate. Do I have to make changes in ‘config_infer_primary.txt’ or anywhere else to make inference efficient?

It depends on your model but not the video. Please refer to DeepStream SDK FAQ - #21 by mchi for the accuracy issues.

Thanks
The intention is to use the pretrained model provided in
deepstream/samples/models/Primary_Detector

I am not so much familiar with training. I wonder if there is any parameter to tune somewhere (in pretrained detector) to make it suitable for very wide image frames (1920x720) too.

@Morganh Can you provide some suggestions for training?

You can set to fp16 or fp32 mode to check if it helps.
For training, end user can use TAO to train their own dataset.

I tried fp16 and fp32, but didn’t find any improvement.

I found some improvements by modifying deepstream/samples/models/Primary_Detector/resnet10.prototxt.

From

layer {
name: “input_1”
type: “Input”
top: “input_1”
input_param {
shape {
dim: 1
dim: 3
dim: 368
dim: 640 } } }

To

layer {
name: “input_1”
type: “Input”
top: “input_1”
input_param {
shape {
dim: 1
dim: 3
dim: 368
dim: 960 } } }

Is it a good approach?

There is no update from you for a period, assuming this is not an issue anymore.
Hence we are closing this topic. If need further support, please open a new one.
Thanks

It may help

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.