/edaps

[ICCV23] EDAPS: Enhanced Domain-Adaptive Panoptic Segmentation

Primary LanguagePythonApache License 2.0Apache-2.0

EDAPS: Enhanced Domain-Adaptive Panoptic Segmentation

by Suman Saha, Lukas Hoyer, Anton Obukhov, Dengxin Dai, and Luc Van Gool

[Arxiv]

🔔 We are happy to announce that EDAPS was accepted at ICCV23. 🔔

Overview

With autonomous industries on the rise, domain adaptation of the visual perception stack is an important research direction due to the cost savings promise. Much prior art was dedicated to domain-adaptive semantic segmentation in the synthetic-to-real context. Despite being a crucial output of the perception stack, panoptic segmentation has been largely overlooked by the domain adaptation community. Therefore, we revisit well-performing domain adaptation strategies from other fields, adapt them to panoptic segmentation, and show that they can effectively enhance panoptic domain adaptation. Further, we study the panoptic network design and propose a novel architecture (EDAPS) designed explicitly for domain-adaptive panoptic segmentation. It uses a shared, domain-robust transformer encoder to facilitate the joint adaptation of semantic and instance features, but task-specific decoders tailored for the specific requirements of both domain-adaptive semantic and instance segmentation. As a result, the performance gap seen in challenging panoptic benchmarks is substantially narrowed. EDAPS significantly improves the state-of-the-art performance for panoptic segmentation UDA by a large margin of 25% on SYNTHIA-to-Cityscapes and even 72% on the more challenging SYNTHIA-to-Mapillary Vistas.

When comparing EDAPS to the previous state-of-the-art UDA method CVRN, we can see that EDAPS has several strengths, as demonstrated in the examples below.

Qualitative Results (SYNTHIA → Cityscapes)

The major improvements come from better panoptic segmentation of the "thing" classes (e.g., "person", "rider", "car") and "stuff" classes (e.g., "traffic light", "traffic sign", "pole") across different object scales, appearance, and viewing angles. In general, EDAPS can better delineate object boundaries, resulting in better-quality pixel-level panoptic segmentation. Note that the detected object shapes predicted by the EDAPS have a better resemblance of real-world objects than the prior art. Thanks to the domain-robust Mix Transformer (MiT-B5) backbone, EDAPS can learn a richer set of domain-invariant semantic and instance features helpful in better segmentation of fine structures.

EDAPS can better segment occluded instances in crowded scenes. Specifically, the "person" segments predicted by EDAPS preserve finer details of the human body even when instances are occluded. Similar observations can be made for the "rider" and "car" classes. For large object instances, such as "bus", EDAPS can segment out the entire object, whereas prior art fails to do so.

Qualitative Results (SYNTHIA → Mapillary Vistas)

In this section, We show visual qualitative results on SYNTHIA → Mapillary Vistas UDA panoptic benchmark.
EDAPS segments better the pole instance. It can be observed that the baseline model struggles to learn the correct class labels and instance masks, whereas EDAPS successfully bridges the domain gap by learning the correct semantics and instances. EDAPS produces better panoptic segmentation for "bus", "rider", "motorbike", "car", "traffic sign".

To learn more about EDAPS, we invite you to refer to our [Paper].

If you find our project valuable for your research, we kindly request that you consider citing it.

@misc{saha2023edaps,
      title={EDAPS: Enhanced Domain-Adaptive Panoptic Segmentation}, 
      author={Suman Saha and Lukas Hoyer and Anton Obukhov and Dengxin Dai and Luc Van Gool},
      year={2023},
      eprint={2304.14291},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Comparison with State of the Art

It can be seen that EDAPS consistently outperforms previous methods with a large margin in all aggregated metrics (mSQ, mRQ, mPQ). Specifically, EDAPS improves the mPQ from 33.0 to 41.2 on SYNTHIA → Cityscapes and from 21.3 to 36.6 SYNTHIA → Mapillary, which is a respective improvement of remarkable 25% and 72% over previous works.

Quantative Results (SYNTHIA → Cityscapes)

mSQ mRQ mPQ
FDA [1] 65.0 35.5 26.6
CRST [2] 60.3 35.6 27.1
AdvEnt [3] 65.6 36.3 28.1
CVRN [4] 66.6 40.9 32.1
UniDAPS [5] 64.7 42.2 33.0
EDAPS (Ours) 72.7 53.6 41.2

Quantative Results (SYNTHIA → Mapillary Vistas)

mSQ mRQ mPQ
FDA [1] 63.8 26.1 19.1
CRST [2] 63.9 25.2 18.8
AdvEnt [3] 63.6 24.7 18.3
CVRN [4] 65.3 28.1 21.3
EDAPS (Ours) 71.7 46.1 36.6
Ensuring Reproducibility

We use the function "set_random_seed" to set the seed for generating random numbers, which allows the users to reproduce the results. However, we noticed a slight non-deterministic behavior in the trained models, i.e., the same model trained with the same seed value multiple times yields slightly different results. We suggest setting the '--deterministic' command line option to "True" to make the training as deterministic as possible. We used the default setting for '--deterministic' (i.e., a "False" value) to train our models.

References:

  1. Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4085–4095, 2020.
  2. Yang Zou, Zhiding Yu, Xiaofeng Liu, BVK Kumar, and Jinsong Wang. Confidence regularized self-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5982–5991, 2019.
  3. Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2517–2526, 2019.
  4. Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Cross-view regularization for domain adaptive panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10133–10144, 2021.
  5. Jingyi Zhang, Jiaxing Huang, and Shijian Lu. Hierarchical mask calibration for unified domain adaptive panoptic segmentation. arXiv preprint arXiv:2206.15083, 2022.

Code Base Features

This repository provides support for several functionalities, including:

  • Dataloading of ground truth labels for both top-down (labels, boxes, and instance masks) and bottom-up (labels, centers, and offsets) panoptic segmentation.
  • Generation of panoptic ground truth labels for the SYNTHIA and Mapillary Vistas datasets following Cityscapes 19 classes format. This allows training and evaluating models on standard UDA benchmarks such as SYNTHIA → Cityscapes and SYNTHIA → Mapillary Vistas.
  • Model evaluation using three different metrics: mIoU (for semantic segmentation), mAP (for instance segmentation), and mPQ (for panoptic segmentation).
  • Evaluation of both top-down and bottom-up based panoptic segmentation models.
  • Visualization of semantic, instance, and panoptic segmentations.

Setup Environment

For this project, we used python 3.8.5. We recommend setting up a new virtual environment:

python -m venv ~/venv/edaps

In this environment, install the required libraries for EDAPS:

Install torch related stuff:

source ~/venv/edaps/bin/activate
pip install torch==1.7.1
pip install torchvision==0.8.2

Install MMCV:

pip install -U openmim
mim install mmcv-full=1.6.1

Install MMDetection:

mkdir -p /path/to/your/codebase/folder
cd /path/to/your/codebase/folder
git clone https://github.com/open-mmlab/mmdetection.git
cd mmdetection
pip install -v -e .

We use MMDetection version 2.25.1 in this work.

Install the other dependecies:

pip install -r requirements.txt

Download Pretrained Weights

MiT-B5 Weights

Download the MiT-B5 weight

MiT-B5 pretrained on ImageNet-1K provided by the official SegFormer repository

Put the pretrained weights in a folder pretrained/ within this project.

For most of the experiments, only mit_b5.pth is necessary.

Checkpoints

Below, we provide checkpoints for EDAPS for different UDA benchmarks.

Extract these checkpoint folders under edaps_experiments/exp-{expId:05d}/checkpoints/. For example, for expId=7, the path would be: edaps_experiments/exp-00007/checkpoints/edaps_synthia_to_cityscapes.

The checkpoints come with the training logs. Please note that:

  • The logs provide the mPQ and mIoU for 19 classes. It is necessary to convert the mPQ and mIoU to the 16 valid classes. Please, refer to this section for converting the mPQ and mIoU.

Setup Datasets

Download Datasets

Synthia: Download SYNTHIA-RAND-CITYSCAPES from here and extract it to data/synthia.

Cityscapes: Download leftImg8bit_trainvaltest.zip and gt_trainvaltest.zip from here and extract them to data/cityscapes.

Mapillary Vistas: Download Mapillary Vistas from here and extract it to data/mapiilary.

Generate Panoptic Ground Truth Labels

One can follow the steps below to generate the ground truth labels for panoptic segmentation. Alternatively, one can download the preprocessed ground truth lables from the below links: here

Synthia:

The creators of the Synthia dataset do not offer panoptic ground truth (GT) labels. Instead, they provide separate semantic and instance labels, which are saved together in a single PNG file. The semantic labels are stored in the first channel of the image, while the instance labels are saved in the second channel. To generate color PNG panoptic GT labels from these files, follow these two steps:

STEP 1:

The following script generates the panoptic GT label files by extracting the semantic and instance GT labels from the two channels of the ground truth PNG files provided with the Synthia dataset. It reads these labels, to create the panoptic GT label files, which are saved in the form of a pickle files.

Run the following command to generate and save the panoptic ground truth labels as pickle files:

./synthiascripts/gen_and_save_panoptic_labels_pickles.sh

One can refer to the Mapillary Vistas ground truth generation script (see below), which leverages the Python multiprocessing library for parallel processing.

STEP 2:

The following script loads the pickle files (generated in Step 1), converts them to RGB PNG files, and saves them to disk.

Run the following command to save the panoptic labels as PNG files:

./synthiascripts/gen_and_save_panoptic_labels_pngs.sh

This script generates the following folder/file:

Folder/File Description
data/synthia/panoptic-labels-crowdth-0-for-daformer/synthia_panoptic This folder contains panoptic ground truth labels for all images of Synthia
data/synthia/panoptic-labels-crowdth-0-for-daformer/synthia_panoptic.json This file contains ground truth information required for training and evaluation

Cityscapes:

To generate the panoptic ground truth labels for Cityscapes, run the following command. This script saves the RGB PNG label files.

Run the following commands for train and validation sets:

./cityscapesscripts/gen_and_save_panoptic_labels_pngs.sh

This script generates the following folder/file:

Folder/File Description
data/cityscapes/gtFine_panoptic/cityscapes_panoptic_train_trainId This folder contains panoptic ground truth labels for train images (needed for Oracle training)
data/cityscapes/gtFine_panoptic/cityscapes_panoptic_train_trainId.json Needed for Oracle training
data/cityscapes/gtFine_panoptic/cityscapes_panoptic_val This folder contains panoptic ground truth labels for validation images (needed for evaluation)
data/cityscapes/gtFine_panoptic/cityscapes_panoptic_val.json Needed for evaluation

Mapillary:

STEP 1:

Mapillary provides the essential data files for training and validating models. These files include the ground truth panoptic PNG images and the corresponding train/validation JSON files, which can be found in the 'train/panoptic/panoptic_2018.json' and 'val/panoptic/panoptic_2018.json' directories. Each ground truth panoptic PNG file has unique panoptic IDs assigned to objects within the image. To create new panoptic ground truth images that are compatible with the Cityscapes 19 classes panoptic segmentation scheme, we load the Mapillary panoptic PNG files and map the original panoptic IDs to the Cityscapes IDs. These new ground truth images are saved as pickle files. The script is designed to run on multiple cores and uses python's multiprocessing to optimize performance. On average, it takes approximately 40 to 50 minutes to process 18,000 training images with 16 cores. Set the DEBUG and VIS flags to False and selecting either the 'train' or 'val' split as needed.

Run the following command to generate and save the panoptic ground truth labels as pickle files:

./mapillaryscripts/gen_and_save_panoptic_labels_pickles.sh

STEP 2:

After generating the new ground truth panoptic images in pickle format, we load the Mapillary labels as numpy arrays from these pickle files. We then convert the numpy array into PNG color images. These color PNG images are saved as the Mapillary ground truth panoptic labels. Additionally, We generate train and validation JSON files used during model training and evaluation.

Run the following commands for the train and val splits:

./mapillaryscripts/gen_and_save_panoptic_labels_pngs.sh

This script generates the following folder/file:

Folder/File Description
data/mapillary/train_panoptic_19cls This folder contains panoptic ground truth labels for the training set (only needed for Oracle training)
data/mapillary/val_panoptic_19cls This folder contains panoptic ground truth labels for validation set (needed for evaluation)
data/mapillary/train_panoptic_19cls.json Only needed for Oracle training
data/mapillary/val_panoptic_19cls.json Needed for evaluation

The final folder structure should look like this:

data
├── cityscapes
│   ├── gtFine
│   │   ├── train
│   │   └── val
│   ├── gtFine_panoptic
│   │   ├── cityscapes_panoptic_train_trainId
│   │   ├── cityscapes_panoptic_train_trainId.json
│   │   ├── cityscapes_panoptic_val
│   │   └── cityscapes_panoptic_val.json
│   ├── leftImg8bit
│   │   ├── train
│   │   └── val
│   ├── train.txt
│   └── val.txt
├── mapillary
│   ├── train_imgs
│   ├── train_panoptic_19cls
│   │   └── train_labels_19cls
│   ├── train_panoptic_19cls.json
│   ├── val_imgs
│   └── val_panoptic_19cls.json
└── synthia
    ├── panoptic-labels-crowdth-0-for-daformer
    │   ├── synthia_panoptic
    │   └── synthia_panoptic.json
    ├── sample_class_stats_dict.json
    ├── sample_class_stats.json
    └── samples_with_class.json

Inference Demo

Already as this point, the provided EDAPS model (pretrained_edaps/pretrained_edaps_weights/latest.pth) can be applied to Cityscapes validation images:

python run_experiments.py --exp 7

This will run inference on Cityscapes validation images using the EDAPS model (trained on Synthia -> Cityscapes).

The predictions (for visualization) will be saved at:

/edaps_experiments/exp-00007/work_dirs/local-exp00007/[experiment-name]/panoptic_eval/[panoptic-eval-folder-name]/visuals/

When judging the predictions, please keep in mind that EDAPS has no access to real-world labels during the training.

Training

Train EDAPS on SYNTHIA to Cityscapes UDA Benchmark using the following command:

python run_experiments.py --config configs/edaps/syn2cs_uda_warm_dfthings_rcs_croppl_a999_edaps_s0.py

In our Git repository, we offer an option to enable debug mode for both training and evaluation purposes. Debug mode is characterized by a batch size of 1, resulting in reduced GPU memory consumption. Additionally, the evaluation is performed on a smaller subset of the Cityscapes validation set, containing only 12 images. This feature serves as a useful tool to quickly test the entire pipeline, including training and evaluation, for correctness. It is especially beneficial when experimenting with new techniques or approaches built on top of EDAPS. To activate debug mode, simply execute the following command for both training and evaluation:

python run_experiments.py --config configs/edaps/syn2cs_uda_warm_dfthings_rcs_croppl_a999_edaps_s0_debug.py

For this you need to download the "gtFine_panoptic_debug.tar" from here, and put the extracted folder under "data/cityscapes".

For other experiments in our paper, we use a system to automatically generate and train the configs:

python run_experimenty.py --exp <ID>

More information about the available experiments and their assigned IDs, can be found in experiments.py and experiments_bottomup.py. The generated configs will be stored in configs/generated/.

Testing & Predictions

Download the EDAPS checkpoint (Synthia→Cityscapes) following the instructions given here. This EDAPS model is trained on Synthia→Cityscapes, and can be tested on the Cityscapes validation set using:

python run_experimenty.py --exp 6

The predictions will be saved for inspection to: edaps_experiments/exp-00006/work_dirs/local-exp00006/[experiment-folder-name]/panoptic_eval/[panoptic-eval-folder-name]/visuals

For example:

  • experiment-folder-name = 230411_1201_syn2cs_dacs_rcs001_cpl_maskrcnn_mitb5_poly10warm_s0_1930c
  • panoptic-eval-folder-name = panop_eval_11-04-2023_12-01-44-325628.

The mPQ, mIoU, mAP and other intermediate evaluation results of the model will be printed to the console. Refer to the end of the evaluation log for more information such as the class-wise PQ, IoU and AP. The evaluation log can be found at: edaps_experiments/exp-00006/work_dirs/local-exp00006/[experiment-folder-name]/[evaluation-log-file-name] For example, evaluation-log-file-name = "evaluation_logs_20230411_121813.log".

When evaluating a model trained on Synthia→Cityscapes or Synthia→Mapillary, please note that the evaluation script calculates the mIoU for all 19 Cityscapes classes. However, Synthia contains only labels for 16 of these classes. Therefore, it is a common practice in UDA to report the mIoU only on these 16 classes. As the PQ and IoU for the 3 missing classes is 0, you can do the conversion:

  • mPQ16 = mPQ19 * 19 / 16.
  • mIoU16 = mIoU19 * 19 / 16.

Framework Structure

This project is based on DAFormer and MMDetection. Please refer to the DAFormer documentation and MMDetection documentation.

The most relevant files for EDAPS are:

Main Scripts:

Data-loading Related Scripts:

Ground Truth Panoptic Label Generation Scripts:

Training and evaluation related scripts:

Acknowledgement:

This project is based on the following open-source projects. We thank their authors for making the source code publically available.

License

This project is released under the Apache License 2.0, while some specific features in this repository are with other licenses. Please refer to LICENSES.md for the careful check, if you are using our code for commercial matters.