An unofficial personal implementation of UM-Adapt, specifically to tackle joint estimation of panoptic segmentation and depth prediction for autonomous driving datasets.

Last update: Nov 25, 2022

Overview

Semisupervised Multitask Learning

This repository is an unofficial and slightly modified implementation of UM-Adapt[1] using PyTorch.

This code primarily deals with the tasks of sematic segmentation, instance segmentation, depth prediction learned in a multi-task setting (with a shared encoder) on a synthetic dataset and then adapted to another dataset with a domain shift. Specifically for this implementation the aim is to learn the three tasks on the Cityscapes Dataset, then adapt and evaluate performance in a fully unsupervised or a semi-supervised setting on the IDD Dataset.

The architecture used for the semantic and instance segmentation model is taken from Panoptic Deeplab[2]. While a choice for the depth decoder is offered between BTS[3] and FCRN-Depth[4].

Usage

The following commands can be used to run the codebase, please make sure to see the respective papers for more details.

To train the base encoder on the Cityscapes (or any other dataset with appropriate modifications) use the following command. Additional flags can also be set as required:

python base_trainer.py --name BaseRun --cityscapes_dir /path/to/cityscapes
Then train the CCR Regularizer as proposed in UM-Adapt with the following command:

python ccr_trainer.py --base_name BaseRun --cityscapes_dir /path/to/cityscapes --hed_path /path/to/pretrained/HED-Network
Unsupervised adaptation to IDD can now be performed using:

python idd_adapter.py --name AdaptIDD --base_name BaseRun --cityscapes_dir /path/to/cityscapes --idd_dir /path/to/idd --hed_path /path/to/pretrained/HED-Network
Further optional semi-supervised fine-tuning can be done using:

python idd_supervised.py --name SupervisedIDD --base_name BaseRun --idd_name AdaptIDD --idd_epoch 10 --idd_dir /path/to/idd --hed_path /path/to/pretrained/HED-Network --supervised_pct 0.5

The code can generally be modified to suit any dataset as required, the base architectures of different decoders as well as the shared encoders can also be altered as needed.

References

If you find this code helpful in your research, please consider citing the following papers.

[1]  @inproceedings{Kundu_2019_ICCV,
        author = {Kundu, Jogendra Nath and Lakkakula, Nishank and Babu, R. Venkatesh},
        title = {UM-Adapt: Unsupervised Multi-Task Adaptation Using Adversarial Cross-Task Distillation},
        booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
        month = {October},
        year = {2019}
    }

[2]  @inproceedings{cheng2020panoptic,
        author={Cheng, Bowen and Collins, Maxwell D and Zhu, Yukun and Liu, Ting and Huang, Thomas S and Adam, Hartwig and Chen, Liang-Chieh},
        title={Panoptic-DeepLab: A Simple, Strong, and Fast Baseline for Bottom-Up Panoptic Segmentation},
        booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
        month = {June},
        year = {2020}
    }

[3]  @article{lee2019big,
        title={From big to small: Multi-scale local planar guidance for monocular depth estimation},
        author={Lee, Jin Han and Han, Myung-Kyu and Ko, Dong Wook and Suh, Il Hong},
        journal={arXiv preprint arXiv:1907.10326},
        year={2019}
}

[4]  @inproceedings{Xie_ICCV_2015,
         author = {Saining Xie and Zhuowen Tu},
         title = {Holistically-Nested Edge Detection},
         booktitle = {IEEE International Conference on Computer Vision},
         year = {2015}
     }

[5]  @misc{pytorch-hed,
         author = {Simon Niklaus},
         title = {A Reimplementation of {HED} Using {PyTorch}},
         year = {2018},
         howpublished = {\url{https://github.com/sniklaus/pytorch-hed}}
    }

If you use either of Cityscapes or IDD datasets, consider citing them

@inproceedings{Cordts2016Cityscapes,
    title={The Cityscapes Dataset for Semantic Urban Scene Understanding},
    author={Cordts, Marius and Omran, Mohamed and Ramos, Sebastian and Rehfeld, Timo and Enzweiler, Markus and Benenson, Rodrigo and Franke, Uwe and Roth, Stefan and Schiele, Bernt},
    booktitle={Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
    year={2016}
}

@article{DBLP:journals/corr/abs-1811-10200,,
    title={IDD: A Dataset for Exploring Problems of Autonomous Navigation in Unconstrained Environments},
    author = {Varma, Girish and Subramanian, Anbumani and Namboodiri, Anoop and Chandraker, Manmohan and Jawahar, C.V.}
    journal={arXiv preprint arXiv:1811.10200},
    year={2018}

Finally, if you use the Xception backbone, please consider citing

@inproceedings{deeplabv3plus2018,
    title={Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation},
    author={Liang-Chieh Chen and Yukun Zhu and George Papandreou and Florian Schroff and Hartwig Adam},
    booktitle={ECCV},
    year={2018}
}

Acknowledgements

Utility functions from many wonderful open-source projects were used, I would like to especially thank the authors of:

An unofficial personal implementation of UM-Adapt, specifically to tackle joint estimation of panoptic segmentation and depth prediction for autonomous driving datasets.

Related tags

Overview

Semisupervised Multitask Learning

Usage

References

Acknowledgements

Owner

Abhinav Atrishi

Code for database and frontend of webpage for Neural Fields in Visual Computing and Beyond.

A collection of inference modules for fastai2

🏃‍♀️ A curated list about human motion capture, analysis and synthesis.

Implementation of CVAE. Trained CVAE on faces from UTKFace Dataset to produce synthetic faces with a given degree of happiness/smileyness.

CDTrans: Cross-domain Transformer for Unsupervised Domain Adaptation

My course projects for the 2021 Spring Machine Learning course at the National Taiwan University (NTU)

ESTDepth: Multi-view Depth Estimation using Epipolar Spatio-Temporal Networks (CVPR 2021)

Official Code for ICML 2021 paper "Revisiting Point Cloud Shape Classification with a Simple and Effective Baseline"

Simulation-based performance analysis of server-less Blockchain-enabled Federated Learning

An implementation of the "Attention is all you need" paper without extra bells and whistles, or difficult syntax

Weakly Supervised Scene Text Detection using Deep Reinforcement Learning

ResNEsts and DenseNEsts: Block-based DNN Models with Improved Representation Guarantees

Multi agent DDPG algorithm written in Python + Pytorch

Semi-automated OpenVINO benchmark_app with variable parameters

Efficient Lottery Ticket Finding: Less Data is More

Compute execution plan: A DAG representation of work that you want to get done. Individual nodes of the DAG could be simple python or shell tasks or complex deeply nested parallel branches or embedded DAGs themselves.

Unified Pre-training for Self-Supervised Learning and Supervised Learning for ASR

Feedback is important: response-aware feedback mechanism for background based conversation

MazeRL is an application oriented Deep Reinforcement Learning (RL) framework

Script for getting information in discord