Estimate the processing speed of several CNN object detection models. Using C++ code, based on the OpenCV and Nvidia GPU. You can choose a GPU or CPU model as you like.

You can test YOLOv3, YOLOv3-tiny, YOLOv4, YOLOv4-tiny, YOLOv7-tiny, and so on from the DarkNet website. Just load the config file and the corresponding weights. All the weights and config files can be found here. weight_file, cfg file.

We have provided the Faster-RCNN-ResNet50 and Faster-RCNN-Inception as the basic models.

The image size is 400x340. We run the detector on a 1080-ti GPU to statistic the processing time of GPU speed. Then, we ran these models on CPU, Intel Core i7-8700k @3.70GHz to test the detection speed of CPU.

You can contact me if you have any questions. (The model and its config file are easy to find; just goole the name of the weights file and config file provided in the c++ code.)

We added some semantic segmentation models into this repo, which includes ENet and FCN-8s and FCN-32s. You can download the corresponding C++ file and run it. You just need to download the model weights and the config files.

1. enet_segmentation.cpp model: model-cityscapes.net
2. fcn_segmentation.cpp config_file: fcn32s-heavy-pascal.prototxt, weights: fcn32s-heavy-pascal.caffemodel