Metadata-Version: 2.1
Name: uotod
Version: 0.3.post2
Summary: Unbalanced Optimal Transport for Object Detection
Author-email: Henri DE PLAEN <henri.deplaen@gmail.com>, Pierre-François DE PLAEN <pierre-francois.deplaen@esat.kuleuven.be>
License:                    GNU LESSER GENERAL PUBLIC LICENSE
                               Version 3, 29 June 2007
        
         Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
         Everyone is permitted to copy and distribute verbatim copies
         of this license document, but changing it is not allowed.
        
        
          This version of the GNU Lesser General Public License incorporates
        the terms and conditions of version 3 of the GNU General Public
        License, supplemented by the additional permissions listed below.
        
          0. Additional Definitions.
        
          As used herein, "this License" refers to version 3 of the GNU Lesser
        General Public License, and the "GNU GPL" refers to version 3 of the GNU
        General Public License.
        
          "The Library" refers to a covered work governed by this License,
        other than an Application or a Combined Work as defined below.
        
          An "Application" is any work that makes use of an interface provided
        by the Library, but which is not otherwise based on the Library.
        Defining a subclass of a class defined by the Library is deemed a mode
        of using an interface provided by the Library.
        
          A "Combined Work" is a work produced by combining or linking an
        Application with the Library.  The particular version of the Library
        with which the Combined Work was made is also called the "Linked
        Version".
        
          The "Minimal Corresponding Source" for a Combined Work means the
        Corresponding Source for the Combined Work, excluding any source code
        for portions of the Combined Work that, considered in isolation, are
        based on the Application, and not on the Linked Version.
        
          The "Corresponding Application Code" for a Combined Work means the
        object code and/or source code for the Application, including any data
        and utility programs needed for reproducing the Combined Work from the
        Application, but excluding the System Libraries of the Combined Work.
        
          1. Exception to Section 3 of the GNU GPL.
        
          You may convey a covered work under sections 3 and 4 of this License
        without being bound by section 3 of the GNU GPL.
        
          2. Conveying Modified Versions.
        
          If you modify a copy of the Library, and, in your modifications, a
        facility refers to a function or data to be supplied by an Application
        that uses the facility (other than as an argument passed when the
        facility is invoked), then you may convey a copy of the modified
        version:
        
           a) under this License, provided that you make a good faith effort to
           ensure that, in the event an Application does not supply the
           function or data, the facility still operates, and performs
           whatever part of its purpose remains meaningful, or
        
           b) under the GNU GPL, with none of the additional permissions of
           this License applicable to that copy.
        
          3. Object Code Incorporating Material from Library Header Files.
        
          The object code form of an Application may incorporate material from
        a header file that is part of the Library.  You may convey such object
        code under terms of your choice, provided that, if the incorporated
        material is not limited to numerical parameters, data structure
        layouts and accessors, or small macros, inline functions and templates
        (ten or fewer lines in length), you do both of the following:
        
           a) Give prominent notice with each copy of the object code that the
           Library is used in it and that the Library and its use are
           covered by this License.
        
           b) Accompany the object code with a copy of the GNU GPL and this license
           document.
        
          4. Combined Works.
        
          You may convey a Combined Work under terms of your choice that,
        taken together, effectively do not restrict modification of the
        portions of the Library contained in the Combined Work and reverse
        engineering for debugging such modifications, if you also do each of
        the following:
        
           a) Give prominent notice with each copy of the Combined Work that
           the Library is used in it and that the Library and its use are
           covered by this License.
        
           b) Accompany the Combined Work with a copy of the GNU GPL and this license
           document.
        
           c) For a Combined Work that displays copyright notices during
           execution, include the copyright notice for the Library among
           these notices, as well as a reference directing the user to the
           copies of the GNU GPL and this license document.
        
           d) Do one of the following:
        
               0) Convey the Minimal Corresponding Source under the terms of this
               License, and the Corresponding Application Code in a form
               suitable for, and under terms that permit, the user to
               recombine or relink the Application with a modified version of
               the Linked Version to produce a modified Combined Work, in the
               manner specified by section 6 of the GNU GPL for conveying
               Corresponding Source.
        
               1) Use a suitable shared library mechanism for linking with the
               Library.  A suitable mechanism is one that (a) uses at run time
               a copy of the Library already present on the user's computer
               system, and (b) will operate properly with a modified version
               of the Library that is interface-compatible with the Linked
               Version.
        
           e) Provide Installation Information, but only if you would otherwise
           be required to provide such information under section 6 of the
           GNU GPL, and only to the extent that such information is
           necessary to install and execute a modified version of the
           Combined Work produced by recombining or relinking the
           Application with a modified version of the Linked Version. (If
           you use option 4d0, the Installation Information must accompany
           the Minimal Corresponding Source and Corresponding Application
           Code. If you use option 4d1, you must provide the Installation
           Information in the manner specified by section 6 of the GNU GPL
           for conveying Corresponding Source.)
        
          5. Combined Libraries.
        
          You may place library facilities that are a work based on the
        Library side by side in a single library together with other library
        facilities that are not Applications and are not covered by this
        License, and convey such a combined library under terms of your
        choice, if you do both of the following:
        
           a) Accompany the combined library with a copy of the same work based
           on the Library, uncombined with any other library facilities,
           conveyed under the terms of this License.
        
           b) Give prominent notice with the combined library that part of it
           is a work based on the Library, and explaining where to find the
           accompanying uncombined form of the same work.
        
          6. Revised Versions of the GNU Lesser General Public License.
        
          The Free Software Foundation may publish revised and/or new versions
        of the GNU Lesser General Public License from time to time. Such new
        versions will be similar in spirit to the present version, but may
        differ in detail to address new problems or concerns.
        
          Each version is given a distinguishing version number. If the
        Library as you received it specifies that a certain numbered version
        of the GNU Lesser General Public License "or any later version"
        applies to it, you have the option of following the terms and
        conditions either of that published version or of any later version
        published by the Free Software Foundation. If the Library as you
        received it does not specify a version number of the GNU Lesser
        General Public License, you may choose any version of the GNU Lesser
        General Public License ever published by the Free Software Foundation.
        
          If the Library as you received it specifies that a proxy can decide
        whether future versions of the GNU Lesser General Public License shall
        apply, that proxy's public statement of acceptance of any version is
        permanent authorization for you to choose that version for the
        Library.
        
Project-URL: Homepage, https://github.com/hdeplaen/uotod
Project-URL: GitHub, https://github.com/hdeplaen/uotod
Project-URL: Documentation, https://uotod.readthedocs.io/en/latest/
Project-URL: Issues, https://github.com/hdeplaen/uotod/issues
Project-URL: E-DUALITY, https://www.esat.kuleuven.be/stadius/E/
Project-URL: ESAT-STADIUS, https://www.esat.kuleuven.be/stadius/
Project-URL: ESAT-PSI, https://www.esat.kuleuven.be/psi/
Keywords: pytorch,machine learning
Classifier: Development Status :: 3 - Alpha
Classifier: Environment :: GPU :: NVIDIA CUDA
Classifier: Intended Audience :: Developers
Classifier: Intended Audience :: Education
Classifier: Intended Audience :: Science/Research
Classifier: Natural Language :: English
Classifier: Operating System :: MacOS
Classifier: Operating System :: Microsoft :: Windows
Classifier: Operating System :: OS Independent
Classifier: Operating System :: POSIX
Classifier: Operating System :: POSIX :: Linux
Classifier: Programming Language :: C++
Classifier: Programming Language :: Python :: 3
Classifier: Programming Language :: Python :: 3.10
Classifier: Programming Language :: Python :: 3.11
Classifier: Programming Language :: Python :: 3.12
Classifier: Programming Language :: Python :: 3.8
Classifier: Programming Language :: Python :: 3.9
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
Classifier: Topic :: Scientific/Engineering :: Information Analysis
Classifier: Topic :: Scientific/Engineering :: Mathematics
Classifier: Topic :: Software Development :: Libraries :: Python Modules
Classifier: Topic :: Utilities
Requires-Python: >=3.8
Description-Content-Type: text/markdown
License-File: LICENSE
Requires-Dist: matplotlib
Requires-Dist: numpy
Requires-Dist: pot
Requires-Dist: torch
Requires-Dist: torchvision

# Unbalanced Optimal Transport: A Unified Framework for Object Detection
<a href="https://hdeplaen.github.io/uotod/" target="_blank">Presentation</a>&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;
<a href="https://openaccess.thecvf.com/content/CVPR2023/papers/De_Plaen_Unbalanced_Optimal_Transport_A_Unified_Framework_for_Object_Detection_CVPR_2023_paper.pdf" target="_blank">Paper</a>&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;
<a href="https://openaccess.thecvf.com/content/CVPR2023/supplemental/De_Plaen_Unbalanced_Optimal_Transport_CVPR_2023_supplemental.pdf" target="_blank">Supplementary</a>&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;
<a href="https://uotod.readthedocs.io/en/latest/" target="_blank">Documentation</a>&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;


![GitHub License](https://img.shields.io/github/license/hdeplaen/uotod)
![PyPI - Downloads](https://img.shields.io/pypi/dm/uotod)
![PyPI - Version](https://img.shields.io/pypi/v/uotod)
[![Documentation Status](https://readthedocs.org/projects/uotod/badge/?version=latest)](https://uotod.readthedocs.io/en/latest/?badge=latest)
[![Test Status](https://github.com/hdeplaen/uotod/actions/workflows/test.yaml/badge.svg?branch=main)](https://github.com/hdeplaen/uotod/actions/workflows/test.yaml)
[![Build Status](https://github.com/hdeplaen/uotod/actions/workflows/build.yaml/badge.svg?branch=main)](https://github.com/hdeplaen/uotod/actions/workflows/build.yaml)

[//]: # (![GitHub all releases]&#40;https://img.shields.io/github/downloads/hdeplaen/uotod/total&#41;)

H. De Plaen, P.-F. De Plaen, J. A. K. Suykens, M. Proesmans, T. Tuytelaars, and L. Van Gool, “Unbalanced Optimal Transport: A Unified Framework for Object Detection,” in *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)*, Jun. 2023, pp. 3198–3207.

This work has be presented at CVPR 2023 in Vancouver, Canada. The paper and additional resources can be found on the [following website](https://hdeplaen.github.io/uotod/). The paper is in open access and can also be found on the [CVF website](https://openaccess.thecvf.com/content/CVPR2023/html/De_Plaen_Unbalanced_Optimal_Transport_A_Unified_Framework_for_Object_Detection_CVPR_2023_paper.html) as well as on [IEEE Xplore](https://ieeexplore.ieee.org/document/10204500).

![Different matching strategies. All are particular cases of Unbalanced Optimal Transport](img/illustration.png)

## Abstract
*TL;DR: We introduce a much more versatile new class of matching strategies unifying many existing ones, as well as being well suited for GPUs.*

During training, supervised object detection tries to correctly match the predicted bounding boxes and associated classification scores to the ground truth. This is essential to determine which predictions are to be pushed towards which solutions, or to be discarded. Popular matching strategies include matching to the closest ground truth box (mostly used in combination with anchors), or matching via the Hungarian algorithm (mostly used in anchor-free methods). Each of these strategies comes with its own properties, underlying losses, and heuristics. We show how Unbalanced Optimal Transport unifies these different approaches and opens a whole continuum of methods in between. This allows for a finer selection of the desired properties. Experimentally, we show that training an object detection model with Unbalanced Optimal Transport is able to reach the state-of-the-art both in terms of Average Precision and Average Recall as well as to provide a faster initial convergence. The approach is well suited for GPU implementation, which proves to be an advantage for large-scale models.

## Install
### PyPI

Using PyPI, it suffices to run `pip install uotod`. Just rerun this command to update the package to its newest version.

### Build From Source

You can also download it directly from the GitHub repository, then build and install it.

```bash
git clone --recursive https://github.com/hdeplaen/uotod
cd uotod
python3 -m pip install -r requirements.txt
python3 -m setup build
python3 -m pip install
 ```

### Compiled Acceleration

The package is **available on all dsitributions** and runs well. However, only the combinations marked with a green ✅ can 
take advantage of the compiled version of Sinkhorn's algorithm directly from PyPI. On a not support combination, you may always build it 
from the source to also have access to Sinkhorn's compiled version of the algorithm. Nevertheless, the PyTorch implementation 
of **Sinkhorn's algorithm is always available** (used by default), this only refers to an additional compiled version. 

| **OS**          	| **Linux** 	| **MacOS** 	 | **Windows** 	|
|-----------------	|:---------:	|:-----------:|:-----------:	|
| **Python 3.8**  	|     ✅     	|   ✅     	   |      ☑️      	|
| **Python 3.9**  	|     ✅     	|   ✅     	   |      ☑️      	|
| **Python 3.10** 	|     ✅     	|   ✅     	   |      ☑️      	|
| **Python 3.11** 	|     ✅     	|   ✅     	   |      ☑️      	|
| **Python 3.12** 	|     ✅     	|     ☑️       |      ☑️      	|

- ✅: Python implementation + compiled acceleration, _both directly from PyPI_
- ☑️: Python implementation _directly from PyPI_ (+ possible compiled acceleration if building from source)

## Examples

### OT matching with GIoU loss:

```python
from uotod.match import BalancedSinkhorn
from uotod.loss import GIoULoss

ot = BalancedSinkhorn(
    loc_match_module=GIoULoss(reduction="none"),
    background_cost=0.,
)
```

### Hungarian matching (bipartite) with GIoU loss:

```python
from uotod.match import Hungarian
from uotod.loss import GIoULoss

hungarian = Hungarian(
    loc_match_module=GIoULoss(reduction="none"),
    background_cost=0.,
)
```

### Loss from SSD solved with Unbalanced Optimal Transport:

```python
from torch.nn import L1Loss, CrossEntropyLoss

from uotod.match import UnbalancedSinkhorn
from uotod.loss import DetectionLoss, IoULoss

matching_method = UnbalancedSinkhorn(
    cls_match_module=None,  # No classification cost
    loc_match_module=IoULoss(reduction="none"),
    background_cost=0.5,  # Threshold for matching to background
    is_anchor_based=True,  # Use anchor-based matching
    reg_target=1e-3,  # Relax the constraint that each ground truth is matched to exactly one prediction
)

loss = DetectionLoss(
    matching_method=matching_method,
    cls_loss_module=CrossEntropyLoss(reduction="none"),
    loc_loss_module=L1Loss(reduction="none"),
)

preds = ...
targets = ...
anchors = ...

loss_value = loss(preds, targets, anchors)
```

### Loss from DETR solved with Optimal Transport (num_classes=3):

```python
import torch
from torch.nn import L1Loss, CrossEntropyLoss

from uotod.match import BalancedSinkhorn
from uotod.loss import DetectionLoss
from uotod.loss import MultipleObjectiveLoss, GIoULoss, NegativeProbLoss

matching_method = BalancedSinkhorn(
    cls_match_module=NegativeProbLoss(reduction="none"),
    loc_match_module=MultipleObjectiveLoss(
        losses=[GIoULoss(reduction="none"), L1Loss(reduction="none")],
        weights=[1., 5.],
    ),
    background_cost=0.,  # Does not influence the matching when using balanced OT
)

loss = DetectionLoss(
    matching_method=matching_method,
    cls_loss_module=CrossEntropyLoss(
        reduction="none",
        weight=torch.tensor([0.1, 1., 1.])  # down-weight the loss for the no-object class
    ),
    loc_loss_module=MultipleObjectiveLoss(
        losses=[GIoULoss(reduction="none"), L1Loss(reduction="none")],
        weights=[1., 5.],
    ),
)

preds = ...
targets = ...
loss_value = loss(preds, targets)
```


## Color Boxes Dataset
![Examples from the Color Boxes Dataset](img/colorboxes.png)

## Citation
If you make any use of our work, please refer to us as:
```bibtex
@InProceedings{De_Plaen_2023_CVPR,
    author    = {De Plaen, Henri and De Plaen, Pierre-Fran\c{c}ois and Suykens, Johan A. K. and Proesmans, Marc and Tuytelaars, Tinne and Van Gool, Luc},
    title     = {Unbalanced Optimal Transport: A Unified Framework for Object Detection},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2023},
    pages     = {3198-3207}
}
```

## Acknowledgements
EU: The research leading to these results has received funding from the European Research Council under the European Union’s Horizon 2020 research and innovation program / ERC Advanced Grant E-DUALITY (787960). This paper reflects only the authors’ views and the Union is not liable for any use that may be made of the contained information. Research Council KUL: Optimization frameworks for deep kernel machines C14/18/068. Flemish Government: FWO: projects: GOA4917N (Deep Restricted Kernel Machines: Methods and Foundations), PhD/Postdoc grant; This research received funding from the Flemish Government (AI Research Program). All the authors are also affiliated to Leuven.AI - KU Leuven institute for AI, B-3000, Leuven, Belgium.
<p style="text-align: center;">
<img src="https://hdeplaen.github.io/uotod/img/eu.png" alt="European Union" style="height:80px;"/>
<img src="https://hdeplaen.github.io/uotod/img/erc.png" alt="European Research Council" style="height:80px;"/>
<img src="https://hdeplaen.github.io/uotod/img/fwo.png" alt="Fonds voor Wetenschappelijk Onderzoek" style="height:80px;"/>
<img src="https://hdeplaen.github.io/uotod/img/vl.png" alt="Flanders AI" style="height:80px;"/>
<img src="https://hdeplaen.github.io/uotod/img/kuleuven.png" alt="KU Leuven" style="height:80px;"/>
</p>
