File size: 8,551 Bytes
9c9498f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
# Gaze-LLE
<div style="text-align:center;">
    <img src="./assets/the_office.png" height="100"/>
    <img src="./assets/MLB_1.gif" height="100"/>
    <img src="./assets/succession.png" height="100"/>
    <img src="./assets/CBS_2.gif" height="100"/>
</div>

[Gaze-LLE: Gaze Target Estimation via Large-Scale Learned Encoders](https://arxiv.org/abs/2412.09586) \
[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](https://github.com/fkryan/gazelle/releases/download/v1.0.0/gazelle_dinov2_vitb14.pt) |
| ```gazelle_dinov2_vitl14``` | DINOv2 ViT-L | ```dinov2_vitl14``` | GazeFollow | [Download](https://github.com/fkryan/gazelle/releases/download/v1.0.0/gazelle_dinov2_vitl14.pt) |
| ```gazelle_dinov2_vitb14_inout``` | DINOv2 ViT-B | ```dinov2_vitb14``` | Gazefollow -> VideoAttentionTarget | [Download](https://github.com/fkryan/gazelle/releases/download/v1.0.0/gazelle_dinov2_vitb14_inout.pt) |
| ```gazelle_large_vitl14_inout``` | DINOv2-ViT-L | ```dinov2_vitl14```  | GazeFollow -> VideoAttentionTarget | [Download](https://github.com/fkryan/gazelle/releases/download/v1.0.0/gazelle_dinov2_vitl14_inout.pt) |


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.

### PyTorch Hub

The models are also available on PyTorch Hub for easy use without installing from source.
```
model, transform = torch.hub.load('fkryan/gazelle', 'gazelle_dinov2_vitb14')
model, transform = torch.hub.load('fkryan/gazelle', 'gazelle_dinov2_vitl14')
model, transform = torch.hub.load('fkryan/gazelle', 'gazelle_dinov2_vitb14_inout')
model, transform = torch.hub.load('fkryan/gazelle', 'gazelle_dinov2_vitl14_inout')
```


## Usage
### Colab Demo Notebook
Check out our [Demo Notebook](https://colab.research.google.com/drive/1TSoyFvNs1-au9kjOZN_fo5ebdzngSPDq?usp=sharing) on Google Colab for how to detect gaze for all people in an image.

### Gaze Prediction
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

```
@article{ryan2024gazelle,
  author       = {Ryan, Fiona and Bati, Ajay and Lee, Sangmin and Bolya, Daniel and Hoffman, Judy and Rehg, James M},
  title        = {Gaze-LLE: Gaze Target Estimation via Large-Scale Learned Encoders},
  journal      = {arXiv preprint arXiv:2412.09586},
  year         = {2024},
}
```

## 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.