Spaces:
Build error
Build error
File size: 9,760 Bytes
a001524 a39ed5d e1c499b 9f5ba58 ad2fc15 9f5ba58 55f4bdb a001524 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 |
import copy
import torch
import numpy as np
import gradio as gr
from spoter_mod.skeleton_extractor import obtain_pose_data
from spoter_mod.normalization.body_normalization import normalize_single_dict as normalize_single_body_dict, BODY_IDENTIFIERS
from spoter_mod.normalization.hand_normalization import normalize_single_dict as normalize_single_hand_dict, HAND_IDENTIFIERS
model = torch.load("spoter-checkpoint.pth", map_location=torch.device('cpu'))
model.train(False)
HAND_IDENTIFIERS = [id + "_Left" for id in HAND_IDENTIFIERS] + [id + "_Right" for id in HAND_IDENTIFIERS]
GLOSS = ['book', 'drink', 'computer', 'before', 'chair', 'go', 'clothes', 'who', 'candy', 'cousin', 'deaf', 'fine',
'help', 'no', 'thin', 'walk', 'year', 'yes', 'all', 'black', 'cool', 'finish', 'hot', 'like', 'many', 'mother',
'now', 'orange', 'table', 'thanksgiving', 'what', 'woman', 'bed', 'blue', 'bowling', 'can', 'dog', 'family',
'fish', 'graduate', 'hat', 'hearing', 'kiss', 'language', 'later', 'man', 'shirt', 'study', 'tall', 'white',
'wrong', 'accident', 'apple', 'bird', 'change', 'color', 'corn', 'cow', 'dance', 'dark', 'doctor', 'eat',
'enjoy', 'forget', 'give', 'last', 'meet', 'pink', 'pizza', 'play', 'school', 'secretary', 'short', 'time',
'want', 'work', 'africa', 'basketball', 'birthday', 'brown', 'but', 'cheat', 'city', 'cook', 'decide', 'full',
'how', 'jacket', 'letter', 'medicine', 'need', 'paint', 'paper', 'pull', 'purple', 'right', 'same', 'son',
'tell', 'thursday']
device = torch.device("cpu")
if torch.cuda.is_available():
device = torch.device("cuda")
def tensor_to_dictionary(landmarks_tensor: torch.Tensor) -> dict:
data_array = landmarks_tensor.numpy()
output = {}
for landmark_index, identifier in enumerate(BODY_IDENTIFIERS + HAND_IDENTIFIERS):
output[identifier] = data_array[:, landmark_index]
return output
def dictionary_to_tensor(landmarks_dict: dict) -> torch.Tensor:
output = np.empty(shape=(len(landmarks_dict["leftEar"]), len(BODY_IDENTIFIERS + HAND_IDENTIFIERS), 2))
for landmark_index, identifier in enumerate(BODY_IDENTIFIERS + HAND_IDENTIFIERS):
output[:, landmark_index, 0] = [frame[0] for frame in landmarks_dict[identifier]]
output[:, landmark_index, 1] = [frame[1] for frame in landmarks_dict[identifier]]
return torch.from_numpy(output)
def greet(label, video0, video1):
if label == "Webcam":
video = video0
elif label == "Video":
video = video1
elif label == "X":
return {"A": 0.8, "B": 0.1, "C": 0.1}
else:
return {}
data = obtain_pose_data(video)
depth_map = np.empty(shape=(len(data.data_hub["nose_X"]), len(BODY_IDENTIFIERS + HAND_IDENTIFIERS), 2))
for index, identifier in enumerate(BODY_IDENTIFIERS + HAND_IDENTIFIERS):
depth_map[:, index, 0] = data.data_hub[identifier + "_X"]
depth_map[:, index, 1] = data.data_hub[identifier + "_Y"]
depth_map = torch.from_numpy(np.copy(depth_map))
depth_map = tensor_to_dictionary(depth_map)
keys = copy.copy(list(depth_map.keys()))
for key in keys:
data = depth_map[key]
del depth_map[key]
depth_map[key.replace("_Left", "_0").replace("_Right", "_1")] = data
depth_map = normalize_single_body_dict(depth_map)
depth_map = normalize_single_hand_dict(depth_map)
keys = copy.copy(list(depth_map.keys()))
for key in keys:
data = depth_map[key]
del depth_map[key]
depth_map[key.replace("_0", "_Left").replace("_1", "_Right")] = data
depth_map = dictionary_to_tensor(depth_map)
depth_map = depth_map - 0.5
inputs = depth_map.squeeze(0).to(device)
outputs = model(inputs).expand(1, -1, -1)
results = torch.nn.functional.softmax(outputs, dim=2).detach().numpy()[0, 0]
results = {GLOSS[i]: float(results[i]) for i in range(100)}
return results
label = gr.outputs.Label(num_top_classes=5, label="Top class probabilities")
demo = gr.Interface(fn=greet, inputs=[gr.Dropdown(["Webcam", "Video"], label="Please select the input type:", type="value"), gr.Video(source="webcam", label="Webcam recording", type="mp4"), gr.Video(source="upload", label="Video upload", type="mp4")], outputs=label,
title="🤟 SPOTER Sign language recognition",
description="""Current user interfaces are not accessible for D/deaf and hard-of-hearing users, whose natural communication medium is sign language. We work on AI systems for sign language to come closer to sign-driven technology and empower accessible apps, websites, and video conferencing platforms.
Try out our recent model for sign language recognition right in your browser! The model below takes a video of a single sign in the American Sign Language at the input and provides you with probabilities of the lemmas (equivalent to words in natural language).
### Our work at CVPR
Our efforts on lightweight and efficient models for sign language recognition were first introduced at WACV with our SPOTER paper. We now presented a work-in-progress follow-up here at CVPR's AVA workshop. Be sure to check our work and code below:
- **WACV2022** - Original SPOTER paper - [Paper](https://openaccess.thecvf.com/content/WACV2022W/HADCV/papers/Bohacek_Sign_Pose-Based_Transformer_for_Word-Level_Sign_Language_Recognition_WACVW_2022_paper.pdf), [Code](https://github.com/matyasbohacek/spoter)
- **CVPR2022 (AVA Worshop)** - Follow-up WIP – [Extended Abstract](https://drive.google.com/file/d/1Szbhi7ZwZ6VAWAcGcDDU6qV9Uj9xnDsS/view?usp=sharing), [Poster](https://drive.google.com/file/d/1_xvmTNbLjTrx6psKdsLkufAtfmI5wfbF/view?usp=sharing)
### How to sign?
The model wrapped in this demo was trained on [WLASL100](https://dxli94.github.io/WLASL/), so it only knows selected ASL vocabulary. Take a look at these tutorial video examples (this is how you sign [computer](https://www.handspeak.com/word/search/index.php?id=449), [work](https://www.handspeak.com/word/search/index.php?id=2423), or [time](https://www.handspeak.com/word/search/index.php?id=2223)), try to replicate them yourself, and have them recognized using the webcam capture below. Have fun!
> The demo can analyze webcam recordings or your uploaded videos. Before you hit Submit, **don't forget to select the input source in the dropdown first**.""",
article="This is joint work of [Matyas Bohacek](https://scholar.google.cz/citations?user=wDy1xBwAAAAJ) and [Zhuo Cao](https://www.linkedin.com/in/zhuo-cao-b0787a1aa/?originalSubdomain=hk). For more info, visit [our website](https://www.signlanguagerecognition.com). To contact us, drop an e-mail [here](mailto:[email protected]).",
css="""
@font-face {
font-family: Graphik;
font-weight: regular;
src: url("https://www.signlanguagerecognition.com/supplementary/GraphikRegular.otf") format("opentype");
}
@font-face {
font-family: Graphik;
font-weight: bold;
src: url("https://www.signlanguagerecognition.com/supplementary/GraphikBold.otf") format("opentype");
}
@font-face {
font-family: MonumentExpanded;
font-weight: regular;
src: url("https://www.signlanguagerecognition.com/supplementary/MonumentExtended-Regular.otf") format("opentype");
}
@font-face {
font-family: MonumentExpanded;
font-weight: bold;
src: url("https://www.signlanguagerecognition.com/supplementary/MonumentExtended-Bold.otf") format("opentype");
}
html {
font-family: "Graphik";
}
h1 {
font-family: "MonumentExpanded";
}
#12 {
- background-image: linear-gradient(to left, #61D836, #6CB346) !important;
background-color: #61D836 !important;
}
#12:hover {
- background-image: linear-gradient(to left, #61D836, #6CB346) !important;
background-color: #6CB346 !important;
border: 0 !important;
border-color: 0 !important;
}
.dark .gr-button-primary {
--tw-gradient-from: #61D836;
--tw-gradient-to: #6CB346;
border: 0 !important;
border-color: 0 !important;
}
.dark .gr-button-primary:hover {
--tw-gradient-from: #64A642;
--tw-gradient-to: #58933B;
border: 0 !important;
border-color: 0 !important;
}
""",
cache_examples=True
)
demo.launch(debug=True)
|