search

smallcorgi / Faster-RCNN_TF

Faster-RCNN in Tensorflow
MIT License
2.34k stars 1.12k forks source link

How to use a pre-trained model to test my own image dataset? #48

Open bitwangdan opened 7 years ago

bitwangdan commented 7 years ago

the pre-trained model is VGGnet_fast_rcnn_iter_70000.ckpt

dakshvar22 commented 7 years ago

Hi, Did you figure this out?

Henrilin28 commented 7 years ago

@dakshvar22 I would like to know how to do it.

ArturoDeza commented 7 years ago

@Henrilin28 @dakshvar22 ; @bitwangdan is correct; The VGG net is the pretrained model on ImageNet (as it says on the README.md). I am currently also figuring out how to retrain on my own dataset. I just started looking into the train_net.py where it's pointing me to check in the /lib/datasets/ folder. You can see that there's multiple dataset files coco.py, kitti.py, etc.. So I'm assuming you'd have to check how they are defining their images and bounding boxes for training.

Hmm, although I have a feeling that this workaround would re-train the entire network from scratch, rather than do some transfer-learning and initialize the weights with the VGG_Net model trained on ImageNet

ghost commented 7 years ago
  1. download vgg16.npy 2.format voc of your datasets 3.according readme to train your dataset 4.very easy