ruizengalways/pfnet

Code Repo of 'Rethinking Planar Homography Estimation Using Perspective Fields'

---

Perspective Field Network

Code Repo of 'Rethinking Planar Homography Estimation Using Perspective Fields'

Tensorflow/Keras implementation for reproducing Perspective Network (PFNet) results in the paper Rethinking Planar Homography Estimation Using Perspective Fields by Rui Zeng, Simon Denman, Sridha Sridharan, Clinton Fookes.

---

COCO Dataset

• Please refer to Common Objects in Context to download the dataset used in the paper.

---

Trained Weights.

• Download our trained weights from provisionally trained weights and put it in the root directory

---

Dependencies

python 3.6

• Tensorflow >= 1.5.0
• Keras >= 2.2.0
• Opencv >= 3.0.0

In addition, please add the project folder to PYTHONPATH and pip install the packages if ImportError: No module named xxx error message occur.

---

Training

• Train a PFNet model on the COCO dataset from scratch:
  • python train.py --dataset=/home/COCO

---

Evaluation

• Evaluate the *.h5 model checkpoint
  • python evaluate.py --dataset=/home/COCO --model=./pfnet.h5

---

Citing PFNet

If you find PFNet useful in your research, please consider citing:

---

@inproceedings{zeng18rethinking,
  author    = {Rui Zeng and Simon Denman and Sridha Sridharan and Clinton Fookes},
  title     = {Rethinking Planar Homography Estimation Using Perspective Fields},
  booktitle = {Asian Conference on Computer Vision (ACCV)},
  year      = {2018},
}

pfnet's Issues

Load weights occupy nearly 8GB GPU memory in evaluation stage

is it normal?

MACE of provided model

Hi @ruizengalways ,

What MACE should the provided model have? I've downloaded pfnet_0200.h5 , run the eval script and ended up with:
MACE Metric: 7.451346

It sounds like a way worse quality than reported in the paper. Could you please help me to reproduce your results?

Best,
Danie;

Training loss is stuck

Training loss seems to be stuck for out of box training with COCO data. Is there any knobs to turn on for the model to start learning ?

Something About branch1

I see when you compute the pf,you use
diff_branch1 but i cant find the diff_branch2
and I can see in your code comment:

    # Two branches. The CNN try to learn the H and inv(H) at the same time. So in the first branch, we just compute the
    #  homography H from the original image to a perturbed image. In the second branch, we just compute the inv(H)

But I cant find the another branch , have you ever try use two branch and got the worse result?

Train custom dataset

Hi @ruizengalways

Thanks for sharing the code. May I know if there is any tutorial showing how to train a custom dataset apart from COCO? Thanks

Weight link die

Hi author,
Your weight link of model is die, can you give me weight file of this model ?
thank you

Identity block

Hi @ruizengalways ,

I've got a question about the identity block. In the paper, on figure 2 the identity block has additional 1x1conv and batch norm on the shortcut connection:

But in the code, there is no 1x1conv + bn in the identity block:

    x = KL.Add()([x, input_tensor])
    x = KL.Activation('relu', name='res' + str(branch) + str(stage) + block + '_out')(x)

So, what's the proper design here?

Best,
Daniel