initial commit

.gitignore

@@ -0,0 +1,166 @@
+# data and checkpoint folders
+data/
+checkpoints/
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/latest/usage/project/#working-with-version-control
+.pdm.toml
+.pdm-python
+.pdm-build/
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 Fiona Ryan
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -0,0 +1,130 @@
+# Gaze-LLE
+<div style="text-align:center;">
+    <img src="./assets/Titanic_1.gif" height="100"/>
+    <img src="./assets/MLB_1.gif" height="100"/>
+    <img src="./assets/CBS_2.gif" height="100"/>
+    <img src="./assets/Sunny_1.gif" height="100"/>
+</div>
+
+[Gaze-LLE: Gaze Target Estimation via Large-Scale Learned Encoders](https://fkryan.github.io/gazelle) \
+[Fiona Ryan](https://fkryan.github.io/), Ajay Bati, [Sangmin Lee](https://sites.google.com/view/sangmin-lee), [Daniel Bolya](https://dbolya.github.io/), [Judy Hoffman](https://faculty.cc.gatech.edu/~judy/)\*, [James M. Rehg](https://rehg.org/)\*
+
+
+This is the official implementation for Gaze-LLE, a transformer approach for estimating gaze targets that leverages the power of pretrained visual foundation models. Gaze-LLE provides a streamlined gaze architecture that learns only a lightweight gaze decoder on top of a frozen, pretrained visual encoder (DINOv2). Gaze-LLE learns 1-2 orders of magnitude fewer parameters than prior works and doesn't require any extra input modalities like depth and pose!
+
+<div style="text-align:center;">
+    <img src="./assets/gazelle_arch.png" height="200"/>
+</div>
+
+
+## Installation
+
+Clone this repo, then create the virtual environment.
+```
+conda env create -f environment.yml
+conda activate gazelle
+pip install -e .
+```
+If your system supports it, consider installing [xformers](https://github.com/facebookresearch/xformers) to speed up attention computation.
+```
+pip3 install -U xformers --index-url https://download.pytorch.org/whl/cu118
+```
+
+## Pretrained Models
+
+We provide the following pretrained models for download.
+| Name | Backbone type | Backbone name | Training data | Checkpoint |
+| ---- | ------------- | ------------- |-------------- | ---------- |
+| ```gazelle_dinov2_vitb14``` | DINOv2 ViT-B | ```dinov2_vitb14 ```| GazeFollow | Download |
+| ```gazelle_dinov2_vitl14``` | DINOv2 ViT-L | ```dinov2_vitl14``` | GazeFollow | Download |
+| ```gazelle_dinov2_vitb14_inout``` | DINOv2 ViT-B | ```dinov2_vitb14 ``` | Gazefollow -> VideoAttentionTarget | Download |
+| ```gazelle_large_vitl14_inout``` | DINOv2-ViT-L | ```dinov2_vitl14```  | GazeFollow -> VideoAttentionTarget | Download |
+
+Note that our Gaze-LLE checkpoints contain only the gaze decoder weights - the DINOv2 backbone weights are downloaded from ```facebookresearch/dinov2``` on PyTorch Hub when the Gaze-LLE model is created in our code.
+
+The GazeFollow-trained models output a spatial heatmap of gaze locations over the scene with values in range ```[0,1]```, where 1 represents the highest probability of the location being a gaze target. The models that are additionally finetuned on VideoAttentionTarget also predict a in/out of frame gaze score in range ```[0,1]``` where 1 represents the person's gaze target being in the frame.
+
+## Usage
+Gaze-LLE is set up for multi-person inference (e.g. for a single image, GazeLLE encodes the scene only once and then uses the features to predict the gaze of multiple people in the image). The input is a batch of image tensors and a list of bounding boxes for each image representing the heads of the people to predict gaze for in each image. The bounding boxes are tuples of form ```(xmin, ymin, xmax, ymax)``` and are in ```[0,1]``` normalized image coordinates. Below we show how to perform inference for a single person in a single image.
+```
+from PIL import Image
+import torch
+from gazelle.model import get_gazelle_model
+
+model, transform = get_gazelle_model("gazelle_dinov2_vitl14_inout")
+model.load_gazelle_state_dict(torch.load("/path/to/checkpoint.pt", weights_only=True))
+model.eval()
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+model.to(device)
+
+image = Image.open("path/to/image.png").convert("RGB")
+input = {
+    "images": transform(image).unsqueeze(dim=0).to(device),    # tensor of shape [1, 3, 448, 448]
+    "bboxes": [[(0.1, 0.2, 0.5, 0.7)]]              # list of lists of bbox tuples
+}
+
+with torch.no_grad():
+    output = model(input)
+predicted_heatmap = output["heatmap"][0][0]        # access prediction for first person in first image. Tensor of size [64, 64]
+predicted_inout = output["inout"][0][0]            # in/out of frame score (1 = in frame) (output["inout"] will be None  for non-inout models)
+```
+We empirically find that Gaze-LLE is effective without a bounding box input for scenes with just one person. However, providing a bounding box can improve results, and is necessary for scenes with multiple people to specify which person's gaze to estimate. To inference without a bounding box, use None in place of a bounding box tuple in the bbox list (e.g. ```input["bboxes"] = [[None]]``` in the example above).
+
+
+We also provide a function to visualize the predicted heatmap for an image.
+```
+import matplotlib.pyplot as plt
+from gazelle.utils import visualize_heatmap
+
+viz = visualize_heatmap(image, predicted_heatmap)
+plt.imshow(viz)
+plt.show()
+```
+
+
+## Evaluate
+We provide evaluation scripts for GazeFollow and VideoAttentionTarget below to reproduce our results from our checkpoints.
+### GazeFollow
+Download the GazeFollow dataset [here](https://github.com/ejcgt/attention-target-detection?tab=readme-ov-file#dataset). We provide a preprocessing script ```data_prep/preprocess_gazefollow.py```, which preprocesses and compiles the annotations into a JSON file for each split within the dataset folder. Run the preprocessing script as
+```
+python data_prep/preprocess_gazefollow.py --data_path /path/to/gazefollow/data_new
+```
+Download the pretrained model checkpoints above and use ```--model_name``` and ```ckpt_path``` to specify the model type and checkpoint for evaluation.
+
+```
+python scripts/eval_gazefollow.py
+    --data_path /path/to/gazefollow/data_new \
+    --model_name gazelle_dinov2_vitl14 \
+    --ckpt_path /path/to/checkpoint.pt \
+    --batch_size 128
+```
+
+
+### VideoAttentionTarget
+Download the VideoAttentionTarget dataset [here](https://github.com/ejcgt/attention-target-detection?tab=readme-ov-file#dataset-1). We provide a preprocessing script ```data_prep/preprocess_vat.py```, which preprocesses and compiles the annotations into a JSON file for each split within the dataset folder. Run the preprocessing script as
+```
+python data_prep/preprocess_gazefollow.py --data_path /path/to/videoattentiontarget
+```
+Download the pretrained model checkpoints above and use ```--model_name``` and ```ckpt_path``` to specify the model type and checkpoint for evaluation.
+```
+python scripts/eval_vat.py
+    --data_path /path/to/videoattentiontarget \
+    --model_name gazelle_dinov2_vitl14_inout \
+    --ckpt_path /path/to/checkpoint.pt \
+    --batch_size 64
+```
+
+## Citation
+
+TBD
+
+## References
+
+- Our models are built on top of pretrained DINOv2 models from PyTorch Hub ([Github repo](https://github.com/facebookresearch/dinov2)).
+
+- Our GazeFollow and VideoAttentionTarget preprocessing code is based on [Detecting Attended Targets in Video](https://github.com/ejcgt/attention-target-detection).
+
+- We use [PyTorch Image Models (timm)](https://github.com/huggingface/pytorch-image-models) for our transformer implementation.
+
+- We use [xFormers](https://github.com/facebookresearch/xformers) for efficient multi-head attention.