Open YashRunwal opened 3 years ago
looks like cspnet use these nets as backbone. they should be trained easier than effnet but much larger, which is not suitable for production. unless they can perform way better than effnet. but i don't have high confidence that it's possible.
Hello,
I have been working on a project which involved detecting objects such as Person, Car, Bicycle, etc. In total there are 6 classes. I read the paper of EfficientDet and liked the idea of BiFPN. But in EffDet "ImageNet-pre-trained EfficientNets" is used as the backbone. However, after that, a new backbone called CSPNet looks promising and I would like to use this backbone along with the BiFPN layer to scale the width, depth, and resolution of the image. The PyTorch implementation of CSPNet is here. I would like to compare the results and draw some conclusions. I will make sure to post the results and more here.
I would like to train the CSPNet+EffDet on the custom image as mentioned above. My images are grayscale images and around 3000 in number. I can build a data loader object for my dataset but I need some help in building the neural network architecture.
I would like some assistance with this as I have not tried this sort of implementation before.
Thank You.