yang7879 / attsets Goto Github PK

View Code? Open in Web Editor NEW

55.0 6.0 20.0 2.53 MB

🔥AttSets in Tensorflow (IJCV 2019)

Home Page: https://doi.org/10.1007/s11263-019-01217-w

License: MIT License

Python 100.00%

multi-view-geometry attention-mechanism 3d-reconstruction 3d-computer-vision deepset 3d-r2n2-dataset

attsets's Introduction

Robust Attentional Aggregation of Deep Feature Sets
for Multi-view 3D Reconstruction

Bo Yang, Sen Wang, Andrew Markham, Niki Trigoni. IJCV, 2019.

(1) Architecture

(2) Optimization

(3) Sample Results

(4) Data

3D-R2N2 Dataset

https://github.com/chrischoy/3D-R2N2

LSM Dataset

https://github.com/akar43/lsm

(5) Released Model

Trained on 3D-R2N2 dataset, 70M

https://drive.google.com/open?id=1A1ihqMDfZLrjQeCFWEjgp-WYb810_om-

(6) Requirements

python 2.7 or 3.5

tensorflow 1.2 +

numpy 1.13.3

scipy 0.19.0

matplotlib 2.0.2

skimage 0.13.0

(7) Run

Training

python main_AttSets.py

Test Demo (Download released model first)

python demo_AttSets.py

(8) Citation

If you use the paper or code for your research, please cite:

@inProceedings{Yang18b,
  title={Robust Attentional Aggregation of Deep Feature Sets for Multi-view 3D Reconstruction},
  author = {Bo Yang
  and Sen Wang
  and Andrew Markham
  and Niki Trigoni,
  booktitle={IJCV 2019},
  year={2018}
}

attsets's People

Contributors

Stargazers

Watchers

attsets's Issues

TypeError: 'float' object cannot be interpreted as an integer

loading files: 03001627
train objs: 4
test objs: 2
X_rgb_train_files_ori: 96
X_rgb_test_files_ori: 48
total_train_batch_num: 1
ep: 0 i: 0 train single rec loss: 0.69298536
Traceback (most recent call last):
File "main_AttSets.py", line 305, in
net.train(data)
File "main_AttSets.py", line 278, in train
X_rgb_batch, Y_vox_batch = data.load_X_Y_test_next_batch(test_mv=1)
File "/Users/ghost/AttSets/tools.py", line 261, in load_X_Y_test_next_batch
idx = random.sample(range(len(self.X_rgb_test_files_ori)/num), self.batch_size)
TypeError: 'float' object cannot be interpreted as an integer

the issue of single-view reconstruction and multi-view reconstruction

I am very inspired by your paper, but I have some questions to ask you.

1.For the experiment of single-view reconstruction :
Category problem: do training set and testing set input per category?
View problem: do training set and testing set input single-view which is random chose?

Similarly, For the experiment of multi-view reconstruction :
Category problem: do training set input multiple categories? do testing set input multiple categories?
View problem: do training set input 24 views? do testing set input multiple views (e.g., 1,2,3,...) ?

2.Could you provide the code of full testing? Does full testing mean to go through all the test set?

How to create a custom data-set and train using it.

Hi, I just experienced pre-trained network and it works fine. But i need to train the network with my own data-set. Please give me a guideline to create a custom data-set for this network.
Also i didn't understood how we can create ground truth. Please help me.

After training testing code gives error like "r2n/Reshape_9:0 is missing

Hi, I have tried retraining of pre-trained network and training from scratch using the training code that you have given. After 400 epoches i have saved the model of both. But i got same error in both model while testing. error was in below line

Y_pred = tf.get_default_graph().get_tensor_by_name("r2n/Reshape_9:0")

r2n/Reshape_9 tensor was not there. I have confirmed it by TensorBoard. But while checking the graph generated by pre-trained network which is downloaded shows the same tensor. How it is possible??

One interesting thing is that , I changed the code line as below

Y_pred = tf.get_default_graph().get_tensor_by_name("r2n/Reshape_7:0")

And then tested with my trained network. It give output. Because after training myself r2n block takes 323232 sized voxels from r2n/Reshape_7:0.

Can you explain why it is so???

About DataLossError

I meet the problem:

DataLossError (see above for traceback): Checksum does not match: stored 769046061 vs. calculated on the restored bytes 1925235876
[[Node: save/RestoreV2_127 = RestoreV2[dtypes=[DT_FLOAT], _device="/job:localhost/replica:0/task:0/device:CPU:0"](_arg_save/Const_0_0, save/RestoreV2_127/tensor_names, save/RestoreV2_127/shape_and_slices)]]
[[Node: save/RestoreV2_112/_225 = _Recvclient_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device_incarnation=1, tensor_name="edge_454_save/RestoreV2_112", tensor_type=DT_FLOAT, _device="/job:localhost/replica:0/task:0/device:GPU:0"]]

Do you know why? My tf version is 1.9.0. THX!