GitPlanet

Cheap and reliable Node.js hosting starts at $3/month, and $1/month static HTML hosting

Created with love in Canada, visit hostnodejs.com today

Feel like to post an Ad? Learn Details
All Projects → RanhaoKang → Pwc Net_pytorch

RanhaoKang / Pwc Net_pytorch

Licence: mit
pytorch implementation of "PWC-Net: CNNs for Optical Flow Using Pyramid, Warping, and Cost Volume"

Programming Languages

python
139335 projects - #7 most used programming language

Labels

pytorchcnnoptical-flow

Projects that are alternatives of or similar to Pwc Net pytorch

Liteflownet2
A Lightweight Optical Flow CNN - Revisiting Data Fidelity and Regularization, TPAMI 2020
Stars: ✭ 195 (+75.68%)
Mutual labels:  cnn, optical-flow
Flownet2 Docker
Dockerfile and runscripts for FlowNet 2.0 (estimation of optical flow)
Stars: ✭ 137 (+23.42%)
Mutual labels:  cnn, optical-flow
Flownet2
FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks
Stars: ✭ 938 (+745.05%)
Mutual labels:  cnn, optical-flow
Hidden Two Stream
Caffe implementation for "Hidden Two-Stream Convolutional Networks for Action Recognition"
Stars: ✭ 179 (+61.26%)
Mutual labels:  cnn, optical-flow
Flownet2 Tf
FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks
Stars: ✭ 378 (+240.54%)
Mutual labels:  cnn, optical-flow
Liteflownet
LiteFlowNet: A Lightweight Convolutional Neural Network for Optical Flow Estimation, CVPR 2018 (Spotlight paper, 6.6%)
Stars: ✭ 474 (+327.03%)
Mutual labels:  cnn, optical-flow
Dispnet Flownet Docker
Dockerfile and runscripts for DispNet and FlowNet1 (estimation of disparity and optical flow)
Stars: ✭ 78 (-29.73%)
Mutual labels:  cnn, optical-flow
Etagger
reference tensorflow code for named entity tagging
Stars: ✭ 100 (-9.91%)
Mutual labels:  cnn
Back2future.pytorch
Unsupervised Learning of Multi-Frame Optical Flow with Occlusions
Stars: ✭ 104 (-6.31%)
Mutual labels:  optical-flow
Facedetector
A re-implementation of mtcnn. Joint training, tutorial and deployment together.
Stars: ✭ 99 (-10.81%)
Mutual labels:  cnn
Captcharecognition
End-to-end variable length Captcha recognition using CNN+RNN+Attention/CTC (pytorch implementation). 端到端的不定长验证码识别
Stars: ✭ 97 (-12.61%)
Mutual labels:  cnn
Lesrcnn
Lightweight Image Super-Resolution with Enhanced CNN (Knowledge-Based Systems,2020)
Stars: ✭ 101 (-9.01%)
Mutual labels:  cnn
Self Driving Car
A End to End CNN Model which predicts the steering wheel angle based on the video/image
Stars: ✭ 106 (-4.5%)
Mutual labels:  cnn
Sarcasmdetection
Sarcasm detection on tweets using neural network
Stars: ✭ 99 (-10.81%)
Mutual labels:  cnn
Places365
The Places365-CNNs for Scene Classification
Stars: ✭ 1,549 (+1295.5%)
Mutual labels:  cnn
Simplepruning
cnn pruning with tensorflow.
Stars: ✭ 98 (-11.71%)
Mutual labels:  cnn
Deeplearning tutorials
The deeplearning algorithms implemented by tensorflow
Stars: ✭ 1,580 (+1323.42%)
Mutual labels:  cnn
Min nlp practice
Chinese & English Cws Pos Ner Entity Recognition implement using CNN bi-directional lstm and crf model with char embedding.基于字向量的CNN池化双向BiLSTM与CRF模型的网络，可能一体化的完成中文和英文分词，词性标注，实体识别。主要包括原始文本数据，数据转换,训练脚本,预训练模型,可用于序列标注研究.注意：唯一需要实现的逻辑是将用户数据转化为序列模型。分词准确率约为93%，词性标注准确率约为90%，实体标注（在本样本上）约为85%。
Stars: ✭ 107 (-3.6%)
Mutual labels:  cnn
Models
DLTK Model Zoo
Stars: ✭ 101 (-9.01%)
Mutual labels:  cnn
Image Regression
🤖️image regression (base on convnetjs)
Stars: ✭ 102 (-8.11%)
Mutual labels:  cnn
View All Similar Projects ➔

Official version(Caffe & PyTorch) is at https://github.com/NVlabs/PWC-Net, thank you all for attention.

News

  • Fix my usage of Correlation Layer, I've been using 19*19 neighborhood for matching.

    NVIDIA is so kind to use their wonderful CUDA to let my mistake seem to be less stupid, btw I don't intend to remove my freaking slow Cost Volume Layer for code diversity or something.

Acknowledgments

PWC-Net

This is an unofficial pytorch implementation of CVPR2018 paper: Deqing Sun et al. "PWC-Net: CNNs for Optical Flow Using Pyramid, Warping, and Cost Volume".
Resources arXiv | Caffe(official)

(flow outputs from top to bottom, the rightest is groundtruth)

It starts to output reasonable flows. However, both time and performance need to be improved. Hope you have fun with this code, and feel free to share your idea about network and its hyper parameters.

Usage

  • Requirements

    • Python 3.6+
    • PyTorch 0.4.0
    • Tensorflow

  • Get Started with Demo
    Note that we only save weights of parameters instead of entire network, provided model file is for default configs, we may upload more advanced models in the future.

    python3 main.py --input_norm --batch_norm --residual --corr Correlation --corr_activation pred --load example/SintelFinal-200K-noBN_SintelFinal-148K-BN.pkl -i example/1.png example/2.png -o example/output.flo

  • Prepare Datasets

    • Download FlyingChairs for training
      filetree when setting --dataset FlyingChairs --dataset_dir <DIR_NAME> 
      <DIR_NAME>
├── 00001_flow.flo
├── 00001_img1.ppm
├── 00001_img2.ppm
...
    • Download FlyingThings for fine-tuning
      filetree when setting --dataset FlyingThings --dataset_dir <DIR_NAME> 
      <DIR_NAME>
    • Download MPI-Sintel for fine-tuning if you want to validate on MPI-Sintel
      filetree when setting --dataset Sintel --dataset_dir <DIR_NAME> 
      <DIR_NAME>
├── training
|   ├── final
|   ├── clean
|   ├── flow
|   ...
├── test
...
    • Download KITTI for fine-tuning if you want to validate on KITTI
      filetree when setting --dataset KITTI --dataset_dir <DIR_NAME> 
      <DIR_NAME>
├── training
|   ├── image_2
|   ├── image_3
|   ...
└── testing

  • Install Correlation Package If you want to use correlation layer (--corr Correlation), please follow NVIDIA/flownet2-pytorch to install extra packages.

  • Train

    python3 main.py train --dataset <DATASET_NAME> --dataset_dir <DIR_NAME>

Details

If there is any difference between your implementation and mine, please create an issue or something.

  • Network Parameters 
    Parameters: 8623340 Size: 32.89543151855469 MB
  • Training Logs 
    Step [100/800000], Loss: 0.3301, EPE: 42.0071, Forward: 34.287192821502686 ms, Backward: 181.38124704360962 ms
Step [200/800000], Loss: 0.2359, EPE: 28.7398, Forward: 32.04517364501953 ms, Backward: 182.32821941375732 ms
Step [300/800000], Loss: 0.2009, EPE: 24.3589, Forward: 31.214130719502766 ms, Backward: 182.9234480857849 ms
Step [400/800000], Loss: 0.1802, EPE: 21.8847, Forward: 31.183505654335022 ms, Backward: 183.74325275421143 ms
Step [500/800000], Loss: 0.1674, EPE: 20.4151, Forward: 30.955915451049805 ms, Backward: 183.9722876548767 ms
Step [600/800000], Loss: 0.1583, EPE: 19.3853, Forward: 30.943967501322426 ms, Backward: 184.35366868972778 ms
Step [700/800000], Loss: 0.1519, EPE: 18.6664, Forward: 30.953510829380583 ms, Backward: 184.56024714878626 ms
Step [800/800000], Loss: 0.1462, EPE: 18.0256, Forward: 30.91249644756317 ms, Backward: 184.76592779159546 ms
Note that the project description data, including the texts, logos, images, and/or trademarks, for each open source project belongs to its rightful owner. If you wish to add or remove any projects, please contact us at [email protected].