Skip to content

GitLab

Commit 754fbc04 authored by bailuo's avatar bailuo
Browse files

init

parent 7aa1ab82
Pipeline #1374 canceled with stages
Hide whitespace changes
Inline Side-by-side
Showing with 957 additions and 0 deletions
LICENSE.txt 0 → 100644
View file @ 754fbc04
Apache License
Version 2.0, January 2004
http://www.apache.org/licenses/
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
1. Definitions.
"License" shall mean the terms and conditions for use, reproduction,
and distribution as defined by Sections 1 through 9 of this document.
"Licensor" shall mean the copyright owner or entity authorized by
the copyright owner that is granting the License.
"Legal Entity" shall mean the union of the acting entity and all
other entities that control, are controlled by, or are under common
control with that entity. For the purposes of this definition,
"control" means (i) the power, direct or indirect, to cause the
direction or management of such entity, whether by contract or
otherwise, or (ii) ownership of fifty percent (50%) or more of the
outstanding shares, or (iii) beneficial ownership of such entity.
"You" (or "Your") shall mean an individual or Legal Entity
exercising permissions granted by this License.
"Source" form shall mean the preferred form for making modifications,
including but not limited to software source code, documentation
source, and configuration files.
"Object" form shall mean any form resulting from mechanical
transformation or translation of a Source form, including but
not limited to compiled object code, generated documentation,
and conversions to other media types.
"Work" shall mean the work of authorship, whether in Source or
Object form, made available under the License, as indicated by a
copyright notice that is included in or attached to the work
(an example is provided in the Appendix below).
"Derivative Works" shall mean any work, whether in Source or Object
form, that is based on (or derived from) the Work and for which the
editorial revisions, annotations, elaborations, or other modifications
represent, as a whole, an original work of authorship. For the purposes
of this License, Derivative Works shall not include works that remain
separable from, or merely link (or bind by name) to the interfaces of,
the Work and Derivative Works thereof.
"Contribution" shall mean any work of authorship, including
the original version of the Work and any modifications or additions
to that Work or Derivative Works thereof, that is intentionally
submitted to Licensor for inclusion in the Work by the copyright owner
or by an individual or Legal Entity authorized to submit on behalf of
the copyright owner. For the purposes of this definition, "submitted"
means any form of electronic, verbal, or written communication sent
to the Licensor or its representatives, including but not limited to
communication on electronic mailing lists, source code control systems,
and issue tracking systems that are managed by, or on behalf of, the
Licensor for the purpose of discussing and improving the Work, but
excluding communication that is conspicuously marked or otherwise
designated in writing by the copyright owner as "Not a Contribution."
"Contributor" shall mean Licensor and any individual or Legal Entity
on behalf of whom a Contribution has been received by Licensor and
subsequently incorporated within the Work.
2. Grant of Copyright License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
copyright license to reproduce, prepare Derivative Works of,
publicly display, publicly perform, sublicense, and distribute the
Work and such Derivative Works in Source or Object form.
3. Grant of Patent License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
(except as stated in this section) patent license to make, have made,
use, offer to sell, sell, import, and otherwise transfer the Work,
where such license applies only to those patent claims licensable
by such Contributor that are necessarily infringed by their
Contribution(s) alone or by combination of their Contribution(s)
with the Work to which such Contribution(s) was submitted. If You
institute patent litigation against any entity (including a
cross-claim or counterclaim in a lawsuit) alleging that the Work
or a Contribution incorporated within the Work constitutes direct
or contributory patent infringement, then any patent licenses
granted to You under this License for that Work shall terminate
as of the date such litigation is filed.
4. Redistribution. You may reproduce and distribute copies of the
Work or Derivative Works thereof in any medium, with or without
modifications, and in Source or Object form, provided that You
meet the following conditions:
(a) You must give any other recipients of the Work or
Derivative Works a copy of this License; and
(b) You must cause any modified files to carry prominent notices
stating that You changed the files; and
(c) You must retain, in the Source form of any Derivative Works
that You distribute, all copyright, patent, trademark, and
attribution notices from the Source form of the Work,
excluding those notices that do not pertain to any part of
the Derivative Works; and
(d) If the Work includes a "NOTICE" text file as part of its
distribution, then any Derivative Works that You distribute must
include a readable copy of the attribution notices contained
within such NOTICE file, excluding those notices that do not
pertain to any part of the Derivative Works, in at least one
of the following places: within a NOTICE text file distributed
as part of the Derivative Works; within the Source form or
documentation, if provided along with the Derivative Works; or,
within a display generated by the Derivative Works, if and
wherever such third-party notices normally appear. The contents
of the NOTICE file are for informational purposes only and
do not modify the License. You may add Your own attribution
notices within Derivative Works that You distribute, alongside
or as an addendum to the NOTICE text from the Work, provided
that such additional attribution notices cannot be construed
as modifying the License.
You may add Your own copyright statement to Your modifications and
may provide additional or different license terms and conditions
for use, reproduction, or distribution of Your modifications, or
for any such Derivative Works as a whole, provided Your use,
reproduction, and distribution of the Work otherwise complies with
the conditions stated in this License.
5. Submission of Contributions. Unless You explicitly state otherwise,
any Contribution intentionally submitted for inclusion in the Work
by You to the Licensor shall be under the terms and conditions of
this License, without any additional terms or conditions.
Notwithstanding the above, nothing herein shall supersede or modify
the terms of any separate license agreement you may have executed
with Licensor regarding such Contributions.
6. Trademarks. This License does not grant permission to use the trade
names, trademarks, service marks, or product names of the Licensor,
except as required for reasonable and customary use in describing the
origin of the Work and reproducing the content of the NOTICE file.
7. Disclaimer of Warranty. Unless required by applicable law or
agreed to in writing, Licensor provides the Work (and each
Contributor provides its Contributions) on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
implied, including, without limitation, any warranties or conditions
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
PARTICULAR PURPOSE. You are solely responsible for determining the
appropriateness of using or redistributing the Work and assume any
risks associated with Your exercise of permissions under this License.
8. Limitation of Liability. In no event and under no legal theory,
whether in tort (including negligence), contract, or otherwise,
unless required by applicable law (such as deliberate and grossly
negligent acts) or agreed to in writing, shall any Contributor be
liable to You for damages, including any direct, indirect, special,
incidental, or consequential damages of any character arising as a
result of this License or out of the use or inability to use the
Work (including but not limited to damages for loss of goodwill,
work stoppage, computer failure or malfunction, or any and all
other commercial damages or losses), even if such Contributor
has been advised of the possibility of such damages.
9. Accepting Warranty or Additional Liability. While redistributing
the Work or Derivative Works thereof, You may choose to offer,
and charge a fee for, acceptance of support, warranty, indemnity,
or other liability obligations and/or rights consistent with this
License. However, in accepting such obligations, You may act only
on Your own behalf and on Your sole responsibility, not on behalf
of any other Contributor, and only if You agree to indemnify,
defend, and hold each Contributor harmless for any liability
incurred by, or claims asserted against, such Contributor by reason
of your accepting any such warranty or additional liability.
END OF TERMS AND CONDITIONS
APPENDIX: How to apply the Apache License to your work.
To apply the Apache License to your work, attach the following
boilerplate notice, with the fields enclosed by brackets "[]"
replaced with your own identifying information. (Don't include
the brackets!) The text should be enclosed in the appropriate
comment syntax for the file format. We also recommend that a
file or class name and description of purpose be included on the
same "printed page" as the copyright notice for easier
identification within third-party archives.
Copyright [yyyy] [name of copyright owner]
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
README.md
View file @ 754fbc04
# OmniMotion
一种在视频序列中密集和长距离运动估计方法，可对运动目标逐像素跟踪。
## 论文
`Tracking Everything Everywhere All at Once`
- https://arxiv.org/abs/2306.05422
## 模型结构
<!-- 此处一句话简要介绍模型结构 -->
先把一个序列表示成一个准3D的规范量，然后通过定义一个双射，这样我们通过一个准3D空间，就可以描述一个完整的运动。
<div align=center>
<img src="./doc/基本原理.png"/>
</div>
## 算法原理
OmniMotion 保留了投影到每个像素的所有场景点的信息，以及它们的相对深度顺序，这让画面中的点即使暂时被遮挡，也能对其进行追踪。将一整个视频序列作为输入, 同时还输入噪声运动估计(例如光流估计), 然后解出一个完整、全局的运动轨迹。然后，添加了一个优化过程，使其可以用任何帧中的任何像素查询表征，以在整个视频中产生平滑、准确的运动轨迹。
<!-- <div align=center>
<img src="./doc/基本原理.png"/>
</div> -->
## 环境配置
```
mv omnimotion_pytoch omnimotion # 去框架名后缀
# -v 路径、docker_name和imageID根据实际情况修改
```
### Docker（方法一）
此处提供[光源](https://www.sourcefind.cn/#/service-details)拉取docker镜像的地址与使用步骤
```
docker pull image.sourcefind.cn:5000/dcu/admin/base/pytorch:1.13.1-ubuntu20.04-dtk23.10-py38 # 本镜像imageID为：0a56ef1842a7
docker run -it -v /path/your_code_data/:/path/your_code_data/ -v /opt/hyhal/:/opt/hyhal/:ro --shm-size=16G --privileged=true --device=/dev/kfd --device=/dev/dri/ --group-add video --name docker_name imageID bash
cd /your_code_path/omnimotion
pip install -r requirements.txt
```
### Dockerfile（方法二）
此处提供dockerfile的使用方法
```
cd /your_code_path/omnimotion/docker
docker build --no-cache -t codestral:latest .
docker run -it -v /path/your_code_data/:/path/your_code_data/ -v /opt/hyhal/:/opt/hyhal/:ro --shm-size=16G --privileged=true --device=/dev/kfd --device=/dev/dri/ --group-add video --name docker_name imageID bash
cd /your_code_path/omnimotion
pip install -r requirements.txt
```
### Anaconda（方法三）
<!-- 此处提供本地配置、编译的详细步骤，例如： -->
关于本项目DCU显卡所需的特殊深度学习库可从[光合](https://developer.hpccube.com/tool/)开发者社区下载安装。
```
DTK驱动：dtk23.10
python：python3.8
pytorch:1.13.1
```
`Tips：以上DTK驱动、python、pytorch等DCU相关工具版本需要严格一一对应`
其它非深度学习库参照requirements.txt安装：
```
pip install -r requirements.txt
```
## 数据集
`DAVIS`
- https://davischallenge.org/index.html
<!-- - 此处填写公开数据集在公司内部的下载地址（数据集存放中心为：[SCNet AIDatasets](http://113.200.138.88:18080/aidatasets) ，模型用到的各公开数据集请分别填上具体地址。），过小权重文件可打包到项目里。 -->
<!-- - 此处填写公开数据集官网下载地址（非必须）。 -->
此处提供数据下载、预处理脚本的使用方法
```
cd /your_code_path/omnimotion/
python get_davis.py # 下载数据集DAVIS-2017-trainval-480p
python main_processing.py # 预处理数据集
```
训练数据目录结构如下，用于正常训练的完整数据集请按此目录结构进行制备：
```
├──DAVIS
├──sequence_1/
├──color/
├──mask/ (optional; only used for visualization purposes)
├──count_maps/
├──features/
├──raft_exhaustive/
├──raft_masks/
├──flow_stats.json
├──sequence_2/
├──...
```
## 训练
<!-- 一般情况下，ModelZoo上的项目提供单机训练的启动方法即可，单机多卡、单机单卡至少提供其一训练方法。 -->
### 单机多卡
```
python train.py --config configs/default.txt # 注意config文件的expname和data_dir参数
```
<!-- ### 单机单卡
```
sh xxx.sh 或python xxx.py
``` -->
## 推理
```
python viz.py --config configs/default.txt
```
## result
<!-- 此处填算法效果测试图（包括输入、输出） -->
训练loss情况，视频序列为`dogs-jump`，绿色为GPU，橘色为DCU
<div align=center>
<img src="./doc/loss.png"/>
</div>
可视化结果
- GPU
<video src="./doc/GPU-dogs-jump_corr_foreground_100000.mp4" controls="controls" width="700" height="200"></video>
- DCU
<video src="./doc/DCU-dogs-jump_corr_foreground_100000.mp4" controls="controls" width="700" height="200"></video>
### 精度
无
<!-- 测试数据：[test data](链接)，使用的加速卡:xxx。
根据测试结果情况填写表格：
| xxx | xxx | xxx | xxx | xxx |
| :------: | :------: | :------: | :------: |:------: |
| xxx | xxx | xxx | xxx | xxx |
| xxx | xx | xxx | xxx | xxx | -->
## 应用场景
### 算法类别
<!-- 参考此分类方法（上传时请去除参考图片），与icon图标类别一致，请勿随意命名： -->
<!-- <div align=center>
<img src="./doc/icon.png"/>
</div> -->
<!-- 超出以上分类的类别命名也可参考此网址中的类别名：https://huggingface.co/ \ -->
`目标跟踪`
### 热点应用行业
<!-- 应用行业的填写需要做大量调研，从而为使用者提供专业、全面的推荐，除特殊算法，通常推荐数量>=3。 -->
`制造,电商,医疗,教育`
<!-- ## 预训练权重 -->
<!-- - 此处填写预训练权重在公司内部的下载地址（预训练权重存放中心为：[SCNet AIModels](http://113.200.138.88:18080/aimodels) ，模型用到的各预训练权重请分别填上具体地址。），过小权重文件可打包到项目里。
- 此处填写公开预训练权重官网下载地址（非必须）。 -->
## 源码仓库及问题反馈
<!-- - 此处填本项目gitlab地址 -->
- https://developer.hpccube.com/codes/bailuo/omnimotion_pytorch
## 参考资料
<!-- - 此处填源github地址（方便使用者查看原github issue）
- 此处填参考项目或教程网址 -->
- https://github.com/qianqianwang68/omnimotion
<!-- `关于model.properties（必要）、LICENSE（必要）、CONTRIBUTORS、模型图标（必要）等其它信息提供参照： `[`ModelZooStd.md`](./ModelZooStd.md)
`各个模型需要保留原项目README.md，改名为README_origin.md即可。` -->
README_o.md 0 → 100644
View file @ 754fbc04
# Tracking Everything Everywhere All at Once
PyTorch Implementation for paper [Tracking Everything Everywhere All at Once]((https://omnimotion.github.io/)), ICCV 2023.
[Qianqian Wang](https://www.cs.cornell.edu/~qqw/) <sup>1,2</sup>,
[Yen-Yu Chang](https://yuyuchang.github.io/) <sup>1</sup>,
[Ruojin Cai](https://www.cs.cornell.edu/~ruojin/) <sup>1</sup>,
[Zhengqi Li](https://zhengqili.github.io/) <sup>2</sup>,
[Bharath Hariharan](https://www.cs.cornell.edu/~bharathh/) <sup>1</sup>,
[Aleksander Holynski](https://holynski.org/) <sup>2,3</sup>,
[Noah Snavely](https://www.cs.cornell.edu/~snavely/) <sup>1,2</sup>
<br>
<sup>1</sup>Cornell University, <sup>2</sup>Google Research, <sup>3</sup>UC Berkeley
#### [Project Page](https://omnimotion.github.io/) | [Paper](https://arxiv.org/pdf/2306.05422.pdf) | [Video](https://www.youtube.com/watch?v=KHoAG3gA024)
## Installation
The code is tested with `python=3.8` and `torch=1.10.0+cu111` on an A100 GPU.
```
git clone --recurse-submodules https://github.com/qianqianwang68/omnimotion/
cd omnimotion/
conda create -n omnimotion python=3.8
conda activate omnimotion
pip install torch==1.10.0+cu111 torchvision==0.11.0+cu111 torchaudio==0.10.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install matplotlib tensorboard scipy opencv-python tqdm tensorboardX configargparse ipdb kornia imageio[ffmpeg]
```
## Training
1. Please refer to the [preprocessing instructions](preprocessing/README.md) for preparing input data
for training OmniMotion. We also provide some processed [data](https://omnimotion.cs.cornell.edu/dataset/)
that you can download, unzip and directly train on. (Note that depending on the network speed,
it may be faster to run the processing script locally than downloading the processed data).
2. With processed input data, run the following command to start training:
```
python train.py --config configs/default.txt --data_dir {sequence_directory}
```
You can view visualizations on tensorboard by running `tensorboard --logdir logs/`.
By default, the script trains 100k iterations which takes 8~9h on an A100 GPU and 12-13h on RTX4090.
If you want to skip the optimization and see what the results/formats look like, we provide the weights
for a few sequences [here](https://drive.google.com/drive/folders/16ekLy-4LTkYAavYrWaKk2qUpJ9TyMXlO?usp=sharing).
You can use `viz.py` to visualize the correspondences produced by the models. Please refer to the next section for more details.
## Visualization
The training pipeline generates visualizations (correspondences, pseudo-depth maps, etc) every certain number of steps (saved in `args.out_dir/vis`).
You can also visualize grid points / trails after training by running:
```
python viz.py --config configs/default.txt --data_dir {sequence_directory}
```
Make sure `expname` and `data_dir` are correctly specified, so that the
model and data can be loaded. By specifying `expname`, the latest checkpoints that match that `expname`
will be loaded. Alternatively, you can specify `ckpt_path` to select a particular checkpoint.
To generate the motion trail visualization, foreground/background segmentation mask is required.
For DAVIS videos one can just use the mask annotations provided by the dataset. For custom videos that don't come with
foreground segmentation masks, you can use [remove.bg](https://www.remove.bg/) to remove the background
for the query frame, download the masked image and set `foreground_mask_path` to its path.
[Here](https://omnimotion.cs.cornell.edu/dataset/mask_0.png) is an example of the masked image for the first frame
of the `butterfly` sequence.
```
python viz.py --config configs/default.txt --data_dir {sequence_directory} --foreground_mask_path {mask_file_path}
```
If you download the provided model weights for a sequence from [here](https://drive.google.com/drive/folders/16ekLy-4LTkYAavYrWaKk2qUpJ9TyMXlO?usp=sharing),
you can visualize the correspondences by running the `viz.py` script and
setting `data_dir` to the unzipped directory, `ckpt_path` to the path for
`model_100000.pth` in the directory, and optionally
`foreground_mask_path`as the path to `mask_0.png`
(only required for non-DAVIS sequences `butterfly`, `kangaroo`, and `swing_tire` if you want to visualize their motion trails).
## Troubleshooting
- The training code utilizes approximately 22GB of CUDA memory. If you encounter CUDA out of memory errors,
you may consider reducing the number of sampled points `num_pts` and the chunk size `chunk_size`.
- Due to the highly non-convex nature of the underlying optimization problem, we observe that the optimization process
can be sensitive to initialization for certain difficult videos. If you notice significant inaccuracies in surface
orderings (by examining the pseudo depth maps) persist after 40k steps,
it is very likely that training won't recover from that. You may consider restarting the training with a
different `loader_seed` to change the initialization.
If surfaces are incorrectly put at the nearest depth planes (which are not supposed to be the closest),
we found using `mask_near` to disable near samples in the beginning of the training could help in some cases.
- Another common failure we noticed is that instead of creating a single object in the canonical space with
correct motion, the method creates duplicated objects in the canonical space with short-ranged motion for each.
This has to do with both that the input correspondences on the object being sparse and short-ranged,
and the optimization being stuck at local minima. This issue may be alleviated with better and longer-range input correspondences
such as from [TAPIR](https://deepmind-tapir.github.io/) and [CoTracker](https://co-tracker.github.io/).
Alternatively, you may consider adjusting `loader_seed` or the learning rates.
## Citation
```
@article{wang2023omnimotion,
title = {Tracking Everything Everywhere All at Once},
author = {Wang, Qianqian and Chang, Yen-Yu and Cai, Ruojin and Li, Zhengqi and Hariharan, Bharath and Holynski, Aleksander and Snavely, Noah},
journal = {ICCV},
year = {2023}
}
```
config.py 0 → 100644
View file @ 754fbc04
import configargparse
def config_parser():
parser = configargparse.ArgumentParser()
parser.add_argument('--config', is_config_file=True, help='config file path')
# general
parser.add_argument('--data_dir', type=str, help='the directory for the video sequence')
parser.add_argument('--expname', type=str, default='', help='experiment name')
parser.add_argument('--local_rank', type=int, default=0, help='rank for distributed training')
parser.add_argument('--save_dir', type=str, default='out/', help='output dir')
parser.add_argument('--ckpt_path', type=str, default='', help='checkpoint path')
parser.add_argument('--no_reload', action='store_true', help='do not reload the weights')
parser.add_argument('--distributed', type=int, default=0, help='if use distributed training')
parser.add_argument('--num_iters', type=int, default=100000, help='number of iterations')
parser.add_argument('--num_workers', type=int, default=4, help='number of workers')
parser.add_argument('--load_opt', type=int, default=1, help='if loading optimizers')
parser.add_argument('--load_scheduler', type=int, default=1, help='if loading schedulers')
parser.add_argument('--loader_seed', type=int, default=12,
help='the random seed used for DataLoader')
# data
parser.add_argument('--dataset_types', type=str, default='flow', help='only flow is included in the current version')
parser.add_argument('--dataset_weights', nargs='+', type=float, default=[1.], help='the weight for each dataset')
parser.add_argument('--num_imgs', type=int, default=250, help='max number of images to train')
parser.add_argument('--num_pairs', type=int, default=8, help='# image pairs to sample in each batch')
parser.add_argument('--num_pts', type=int, default=256, help='# pts to sample from each pair of images')
# lr
parser.add_argument('--lr_feature', type=float, default=1e-3, help='learning rate for feature mlp')
parser.add_argument('--lr_deform', type=float, default=1e-4, help='learning rate for deform mlp')
parser.add_argument('--lr_color', type=float, default=3e-4, help='learning rate for color mlp')
parser.add_argument("--lrate_decay_steps", type=int, default=20000,
help='decay learning rate by a factor every specified number of steps')
parser.add_argument("--lrate_decay_factor", type=float, default=0.5,
help='decay learning rate by a factor every specified number of steps')
parser.add_argument("--grad_clip", type=float, default=0, help='clip the gradient to avoid training instability')
# network training
parser.add_argument('--use_error_map', action='store_true', help='use error map')
parser.add_argument('--use_count_map', action='store_true', help='use count map')
parser.add_argument('--use_affine', action='store_true',
help='if using additional 2D affine transformation layers for x, y in the invertible network')
parser.add_argument('--mask_near', action='store_true',
help='if mask out the nearest samples in the beginning of the optimization,'
'may be helpful to avoid bad initialization associated with wrong surface ordering'
'e.g., a surface is initialized at very small depth but should instead be farther away')
parser.add_argument('--num_samples_ray', type=int, default=32, help='number of samples per ray')
parser.add_argument('--pe_freq', type=int, default=4, help='the freq for pe used in the affine coupling layers')
parser.add_argument('--min_depth', type=float, default=0, help='the minimum depth value')
parser.add_argument('--max_depth', type=float, default=2, help='the maximum depth value')
parser.add_argument('--start_interval', type=int, default=20, help='the starting interval')
parser.add_argument('--max_padding', type=float, default=0,
help='if predicted pixel locs exceed this padding, mask them out for training')
# inference
parser.add_argument('--chunk_size', type=int, default=1000, help='chunk size for rendering depth and rgb')
parser.add_argument('--use_max_loc', action='store_true',
help='during inference, if using only the sample with maximum blending weight on the ray'
'to compute correspondence. If set to False, the correspondences will be computed'
'the same way as training, i.e., compositing all samples along the ray.')
parser.add_argument('--query_frame_id', type=int, default=0, help='the id of the query frame')
parser.add_argument('--vis_occlusion', action='store_true',
help='if marking occluded pixels as crosses for visualization')
parser.add_argument('--occlusion_th', type=float, default=0.99,
help='to determine if a mapped 3d location in the target frame is occluded or not,'
' we look at the fraction of light absorbed by samples in front of this location '
'on the ray in the target frame (i.e., 1 - transmittance)'
'if that value is higher than this threshold, the mapped point is considered as occluded')
parser.add_argument('--foreground_mask_path', type=str, default='',
help='providing the path for foreground mask file for generating trails')
# log
parser.add_argument('--i_print', type=int, default=100, help='frequency for printing losses')
parser.add_argument('--i_img', type=int, default=500, help='frequency for writing visualizations to tensorboard')
parser.add_argument('--i_weight', type=int, default=20000, help='frequency for saving ckpts')
parser.add_argument('--i_cache', type=int, default=20000, help='frequency for caching current flow predictions')
parser.add_argument("-f", "--fff", help="a dummy argument to fool ipython", default="1")
args = parser.parse_args()
return args
configs/default.txt 0 → 100644
View file @ 754fbc04
expname = DCU-xxx
data_dir = /your_code_path/omnimotion/DAVIS_data_path/xxx # 指定一个视频序列
# training
num_pairs = 8
num_pts = 256
use_affine = True
use_error_map = True
use_count_map = True
# inference
use_max_loc = True
vis_occlusion = True
\ No newline at end of file
criterion.py 0 → 100644
View file @ 754fbc04
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision import models
import ssl
ssl._create_default_https_context = ssl._create_unverified_context
def cauchy_loss(pred, gt, c=1, mask=None, normalize=True):
loss = torch.log(1 + ((pred - gt) / c)**2)
if mask is not None:
if normalize:
return (loss * mask).mean() / (mask.mean() + 1e-8)
else:
return (loss * mask).mean()
else:
return loss.mean()
def masked_mse_loss(pred, gt, mask=None, normalize=True):
if mask is None:
return F.mse_loss(pred, gt)
else:
sum_loss = F.mse_loss(pred, gt, reduction='none')
ndim = sum_loss.shape[-1]
if normalize:
return torch.sum(sum_loss * mask) / (ndim * torch.sum(mask) + 1e-8)
else:
return torch.mean(sum_loss * mask)
def masked_l1_loss(pred, gt, mask=None, normalize=True, quantile=1):
if mask is None:
return trimmed_l1_loss(pred, gt, quantile)
else:
sum_loss = F.l1_loss(pred, gt, reduction='none').mean(dim=-1, keepdim=True)
loss_at_quantile = torch.quantile(sum_loss, quantile)
quantile_mask = (sum_loss < loss_at_quantile).squeeze(-1)
ndim = sum_loss.shape[-1]
if normalize:
return torch.sum((sum_loss * mask)[quantile_mask]) / (ndim * torch.sum(mask[quantile_mask]) + 1e-8)
else:
return torch.mean((sum_loss * mask)[quantile_mask])
def masked_huber_loss(pred, gt, delta, mask=None, normalize=True):
if mask is None:
return F.huber_loss(pred, gt, delta=delta)
else:
sum_loss = F.huber_loss(pred, gt, delta=delta, reduction='none')
ndim = sum_loss.shape[-1]
if normalize:
return torch.sum(sum_loss * mask) / (ndim * torch.sum(mask) + 1e-8)
else:
return torch.mean(sum_loss * mask)
def trimmed_l1_loss(pred, gt, quantile=0.9):
loss = F.l1_loss(pred, gt, reduction='none').mean(dim=-1)
loss_at_quantile = torch.quantile(loss, quantile)
trimmed_loss = loss[loss < loss_at_quantile].mean()
return trimmed_loss
def trimmed_std_normed_l1_loss(pred, gt, quantile=0.9):
loss = F.l1_loss(pred, gt, reduction='none') # [..., d]
mask = loss.mean(dim=-1) < torch.quantile(loss.mean(dim=-1), quantile) # [...]
pred_std = torch.std(pred[mask], dim=0) # [d]
gt_std = torch.std(gt[mask], dim=0) # [d]
std = 0.5 * (pred_std + gt_std)
trimmed_std_normed_loss = (loss / std).mean()
return trimmed_std_normed_loss
def trimmed_mse_loss(pred, gt, mask=None, quantile=0.9):
loss = F.mse_loss(pred, gt, reduction='none').mean(dim=-1)
loss_at_quantile = torch.quantile(loss, quantile)
trimmed_loss = loss[loss < loss_at_quantile]
if mask is not None:
mask = mask[loss < loss_at_quantile]
loss = torch.mean(mask * trimmed_loss) / torch.mean(mask)
else:
loss = torch.mean(trimmed_loss)
return loss
def trimmed_var_normed_mse_loss(pred, gt, quantile=0.9):
loss = F.mse_loss(pred, gt, reduction='none') # [..., d]
mask = loss.mean(dim=-1) < torch.quantile(loss.mean(dim=-1), quantile) # [...]
pred_var = torch.var(pred[mask], dim=0) # [d]
gt_var = torch.var(gt[mask], dim=0) # [d]
var = 0.5 * (pred_var + gt_var)
trimmed_var_normed_loss = (loss / var).mean()
return trimmed_var_normed_loss
def compute_depth_range_loss(depth, min_th=0, max_th=2):
'''
the depth of mapped 3d locations should also be within the near and far depth range
'''
loss_lower = ((depth[depth < min_th] - min_th)**2).sum() / depth.numel()
loss_upper = ((depth[depth > max_th] - max_th)**2).sum() / depth.numel()
return loss_upper + loss_lower
def lossfun_distortion(t, w):
"""Compute iint w[i] w[j] |t[i] - t[j]| di dj."""
# The loss incurred between all pairs of intervals.
ut = (t[..., 1:] + t[..., :-1]) / 2
dut = torch.abs(ut[..., :, None] - ut[..., None, :])
loss_inter = torch.sum(w * torch.sum(w[..., None, :] * dut, dim=-1), dim=-1)
# The loss incurred within each individual interval with itself.
loss_intra = torch.sum(w**2 * (t[..., 1:] - t[..., :-1]), dim=-1) / 3
return (loss_inter + loss_intra).mean()
def median_scale_shift(x):
'''
:param x: [batch, h, w]
:return: median scaled and shifted x
'''
batch_size = len(x)
median_x = torch.median(x.reshape(batch_size, -1), dim=1).values[:, None, None]
s_x = torch.mean(torch.abs(x - median_x), dim=(1, 2), keepdim=True)
return (x - median_x) / s_x
def scale_shift_invariant_loss(pred, gt):
pred_ = median_scale_shift(pred)
gt_ = median_scale_shift(gt)
return torch.mean(torch.abs(pred_ - gt_))
def trimmed_scale_shift_invariant_loss(pred, gt, percentile=0.8):
pred_ = median_scale_shift(pred)
gt_ = median_scale_shift(gt)
error = torch.abs(pred_ - gt_).flatten()
cut_value = torch.quantile(error, percentile)
return error[error < cut_value].mean()
class GANLoss(nn.Module):
def __init__(self, gan_mode, target_real_label=1.0, target_fake_label=0.0,
tensor=torch.FloatTensor, opt=None):
super(GANLoss, self).__init__()
self.real_label = target_real_label
self.fake_label = target_fake_label
self.real_label_tensor = None
self.fake_label_tensor = None
self.zero_tensor = None
self.Tensor = tensor
self.gan_mode = gan_mode
self.opt = opt
if gan_mode == 'ls':
pass
elif gan_mode == 'original':
pass
elif gan_mode == 'w':
pass
elif gan_mode == 'hinge':
pass
else:
raise ValueError('Unexpected gan_mode {}'.format(gan_mode))
def get_target_tensor(self, input, target_is_real):
if target_is_real:
if self.real_label_tensor is None:
self.real_label_tensor = self.Tensor(1).fill_(self.real_label)
self.real_label_tensor.requires_grad_(False)
return self.real_label_tensor.expand_as(input)
else:
if self.fake_label_tensor is None:
self.fake_label_tensor = self.Tensor(1).fill_(self.fake_label)
self.fake_label_tensor.requires_grad_(False)
return self.fake_label_tensor.expand_as(input)
def get_zero_tensor(self, input):
if self.zero_tensor is None:
self.zero_tensor = self.Tensor(1).fill_(0)
self.zero_tensor.requires_grad_(False)
return self.zero_tensor.expand_as(input)
def loss(self, input, target_is_real, for_discriminator=True):
if self.gan_mode == 'original': # cross entropy loss
target_tensor = self.get_target_tensor(input, target_is_real)
loss = F.binary_cross_entropy_with_logits(input, target_tensor)
return loss
elif self.gan_mode == 'ls':
target_tensor = self.get_target_tensor(input, target_is_real)
return F.mse_loss(input, target_tensor)
elif self.gan_mode == 'hinge':
if for_discriminator:
if target_is_real:
minval = torch.min(input - 1, self.get_zero_tensor(input))
loss = -torch.mean(minval)
else:
minval = torch.min(-input - 1, self.get_zero_tensor(input))
loss = -torch.mean(minval)
else:
assert target_is_real, "The generator's hinge loss must be aiming for real"
loss = -torch.mean(input)
return loss
else:
# wgan
if target_is_real:
return -input.mean()
else:
return input.mean()
def __call__(self, input, target_is_real, for_discriminator=True):
# computing loss is a bit complicated because |input| may not be
# a tensor, but list of tensors in case of multiscale discriminator
if isinstance(input, list):
loss = 0
for pred_i in input:
if isinstance(pred_i, list):
pred_i = pred_i[-1]
loss_tensor = self.loss(pred_i, target_is_real, for_discriminator)
bs = 1 if len(loss_tensor.size()) == 0 else loss_tensor.size(0)
new_loss = torch.mean(loss_tensor.view(bs, -1), dim=1)
loss += new_loss
return loss / len(input)
else:
return self.loss(input, target_is_real, for_discriminator)
class Vgg16(nn.Module):
def __init__(self):
super(Vgg16, self).__init__()
features = models.vgg16(pretrained=True).features
self.to_relu_1_2 = nn.Sequential()
self.to_relu_2_2 = nn.Sequential()
self.to_relu_3_3 = nn.Sequential()
self.to_relu_4_3 = nn.Sequential()
for x in range(4):
self.to_relu_1_2.add_module(str(x), features[x])
for x in range(4, 9):
self.to_relu_2_2.add_module(str(x), features[x])
for x in range(9, 16):
self.to_relu_3_3.add_module(str(x), features[x])
for x in range(16, 23):
self.to_relu_4_3.add_module(str(x), features[x])
# don't need the gradients, just want the features
for param in self.parameters():
param.requires_grad = False
def forward(self, x):
h = self.to_relu_1_2(x)
h_relu_1_2 = h
h = self.to_relu_2_2(h)
h_relu_2_2 = h
h = self.to_relu_3_3(h)
h_relu_3_3 = h
h = self.to_relu_4_3(h)
h_relu_4_3 = h
out = [h_relu_1_2, h_relu_2_2, h_relu_3_3, h_relu_4_3]
return out
class Vgg19(nn.Module):
def __init__(self, requires_grad=False):
super(Vgg19, self).__init__()
vgg_pretrained_features = models.vgg19(pretrained=True).features
self.slice1 = torch.nn.Sequential()
self.slice2 = torch.nn.Sequential()
self.slice3 = torch.nn.Sequential()
self.slice4 = torch.nn.Sequential()
self.slice5 = torch.nn.Sequential()
for x in range(2):
self.slice1.add_module(str(x), vgg_pretrained_features[x])
for x in range(2, 7):
self.slice2.add_module(str(x), vgg_pretrained_features[x])
for x in range(7, 12):
self.slice3.add_module(str(x), vgg_pretrained_features[x])
for x in range(12, 21):
self.slice4.add_module(str(x), vgg_pretrained_features[x])
for x in range(21, 30):
self.slice5.add_module(str(x), vgg_pretrained_features[x])
if not requires_grad:
for param in self.parameters():
param.requires_grad = False
def forward(self, x):
h_relu1 = self.slice1(x)
h_relu2 = self.slice2(h_relu1)
h_relu3 = self.slice3(h_relu2)
h_relu4 = self.slice4(h_relu3)
h_relu5 = self.slice5(h_relu4)
out = [h_relu1, h_relu2, h_relu3, h_relu4, h_relu5]
return out
class VGGLoss(nn.Module):
def __init__(self, model='vgg19', device='cuda'):
super().__init__()
if model == 'vgg16':
self.vgg = Vgg16().to(device)
self.weights = [1.0/16, 1.0/8, 1.0/4, 1.0]
elif model == 'vgg19':
self.vgg = Vgg19().to(device)
self.weights = [1.0/32, 1.0/16, 1.0/8, 1.0/4, 1.0]
# self.weights = [1/2.6, 1/4.8, 1/3.7, 1/5.6, 10/1.5]
# self.weights = [1/2.6, 1/4.8, 1/3.7, 1/5.6, 2/1.5]
# self.criterion = nn.L1Loss()
self.loss_func = masked_l1_loss
@staticmethod
def preprocess(x, size=224):
# B, C, H, W
min_in_size = min(x.shape[-2:])
device = x.device
mean = torch.tensor([0.485, 0.456, 0.406]).to(device)
std = torch.tensor([0.229, 0.224, 0.225]).to(device)
x = (x - mean.reshape(1, 3, 1, 1)) / std.reshape(1, 3, 1, 1)
# if min_in_size <= size:
# mode = 'bilinear'
# align_corners = True
# else:
# mode = 'area'
# align_corners = None
# x = F.interpolate(x, size=size, mode=mode, align_corners=align_corners)
return x
def forward(self, x, y, mask=None, size=224):
x = self.preprocess(x, size=size) # assume x, y are inside (0, 1)
y = self.preprocess(y, size=size)
if mask is not None:
if min(mask.shape[-2:]) <= size:
mode = 'bilinear'
align_corners = True
else:
mode = 'area'
align_corners = None
mask = F.interpolate(mask, size=size, mode=mode, align_corners=align_corners)
x_vgg, y_vgg = self.vgg(x), self.vgg(y)
# loss = 0
loss = self.loss_func(x, y, mask)
for i in range(len(x_vgg)):
loss += self.weights[i] * self.loss_func(x_vgg[i], y_vgg[i], mask)
return loss
def normalize_minus_one_to_one(x):
x_min = x.min()
x_max = x.max()
return 2. * (x - x_min) / (x_max - x_min) - 1.
def get_flow_smoothness_loss(flow, alpha):
flow_gradient_x = flow[:, :, :, 1:, :] - flow[:, :, :, -1:, :]
flow_gradient_y = flow[:, :, :, :, 1:] - flow[:, :, :, :, -1:]
cost_x = (alpha[:, :, :, 1:, :] * torch.norm(flow_gradient_x, dim=2, keepdim=True)).sum()
cost_y = (alpha[:, :, :, :, 1:] * torch.norm(flow_gradient_y, dim=2, keepdim=True)).sum()
avg_cost = (cost_x + cost_y) / (2 * alpha.sum() + 1e-6)
return avg_cost
dataset/Davis.txt 0 → 100644
View file @ 754fbc04
数据集davis，参考preprocessing/README.md
\ No newline at end of file
docker/Dockerfile 0 → 100644
View file @ 754fbc04
FROM image.sourcefind.cn:5000/dcu/admin/base/pytorch:2.1.0-ubuntu20.04-dtk24.04.1-py3.10
ENV DEBIAN_FRONTEND=noninteractive
# COPY requirements.txt requirements.txt
# RUN pip3 install -r requirements.txt -i http://mirrors.aliyun.com/pypi/simple/ --trusted-host mirrors.aliyun.com
get_davis.py 0 → 100644
View file @ 754fbc04
# run: python get_davis.py <OUT_DIR>
# this file converts the DAVIS dataset into our format.
import os
import shutil
import sys
import subprocess
subprocess.run(['wget', 'https://data.vision.ee.ethz.ch/csergi/share/davis/DAVIS-2017-trainval-480p.zip'])
subprocess.run(['unzip', 'DAVIS-2017-trainval-480p.zip'])
img_src_root = 'DAVIS/JPEGImages/480p/'
seq_names = os.listdir(img_src_root)
out_dir = sys.argv[1]
os.makedirs(out_dir, exist_ok=True)
for seq_name in seq_names:
img_src_dir = os.path.join(img_src_root, seq_name)
img_dst_dir = os.path.join(out_dir, seq_name, 'color')
shutil.copytree(img_src_dir, img_dst_dir)
# mask is used only for visualization purposes
mask_src_root = 'DAVIS/Annotations/480p/'
mask_src_dir = os.path.join(mask_src_root, seq_name)
mask_dst_dir = os.path.join(out_dir, seq_name, 'mask')
shutil.copytree(mask_src_dir, mask_dst_dir)
print('DAVIS data is saved to: {}'.format(os.path.abspath(out_dir)))
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment