Marcos Luciano comments

Results 468 comments of


                                            Marcos Luciano

yolov7 segmentation model support

I have no plans for segmentation for now.

Do Inference size width and height have to be equal?

It doesn't need to be width=height but you must use the same size of the trained model (640) to get the correct output. The sizes must be multiple of 32.

replicating the results in readme.md

I got the results from a custom code that I created. It's not public for now.

replicating the results in readme.md

> Okay, I am have written code to check the performance using deepstream. You can just use the deepstream-app to check the FPS. > I got, I find that yolov7...

Too low fps at T4 and V100.

Please attach the config_infer file and deepstream_app_config.txt file.

Too low fps at T4 and V100.

You are using too many resources in your `rtsp_analytics_video.txt` config. Probably, this is the problem.

Too low fps at T4 and V100.

You need to create a custom code (preferable in C) to optimize the pipeline for what you want to do. The `deepstream-app` is like a general purpose application.

Too low fps at T4 and V100.

What's the framerate when you use default `deepstream_app_config.txt` file?

YOLOV5 model has very weak detection capability after deployment

> `Jetpack 4.6 deepstream 6.0.1 yolov5 6.1 ` I tested yolov5s , yolov7 , yolor （untrained,all of them），the can work well. I trained yolov5（changed backbone，but layers(CBH , LC_block , Dense)...

Is batch_nms basically included in the TensorRT build?

I reverted the NMS to DeepStream default NMS due to accurary drop, so the Batched NMS isn't used anymore.