# Copyright 2022 AI Singapore
#
# 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
#
# https://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.
"""👨👩👧👦 Congested Scene Recognition network: Dilated convolutional neural
networks for understanding the highly congested scenes.
"""
from typing import Any, Dict, Optional
import cv2
from peekingduck.pipeline.nodes.abstract_node import AbstractNode
from peekingduck.pipeline.nodes.model.csrnetv1 import csrnet_model
[docs]class Node(AbstractNode):
"""Initializes and uses CSRNet model to predict the density map and crowd
count.
The csrnet node is capable of predicting the number of people in dense and
sparse crowds. The dense and sparse crowd models were trained using data
from ShanghaiTech Part A and ShanghaiTech Part B respectively. As the
models were trained to recognize congested scenes, the estimates are less
accurate if the number of people are low (e.g. less than 10).
Inputs:
|img_data|
Outputs:
|density_map_data|
|count_data|
Configs:
model_type (:obj:`str`): **{"dense", "sparse"}, default="sparse"**. |br|
Defines the type of CSRNet model to be used. The node uses the
sparse crowd model by default and can be changed to using the dense
crowd model. As a rule of thumb, the dense crowd model should be
used if the people in a given image or video frame are packed
shoulder to shoulder, e.g., stadiums.
weights_parent_dir (:obj:`Optional[str]`): **default = null**. |br|
Change the parent directory where weights will be stored by
replacing ``null`` with an absolute path to the desired directory.
width (:obj:`int`): **default = 640**. |br|
By default, the width of an image will be resized to 640 for
inference. The height of the image will be resized proportionally
to preserve its aspect ratio. In general, decreasing the width of
an image will improve inference speed. However, this might impact
the accuracy of the model.
References:
CSRNet: Dilated Convolutional Neural Networks for Understanding the
Highly Congested Scenes:
https://arxiv.org/pdf/1802.10062.pdf
Model weights trained by https://github.com/Neerajj9/CSRNet-keras
Inference code adapted from https://github.com/Neerajj9/CSRNet-keras
"""
def __init__(self, config: Dict[str, Any] = None, **kwargs: Any) -> None:
super().__init__(config, node_path=__name__, **kwargs)
self.model = csrnet_model.CSRNetModel(self.config)
def run(self, inputs: Dict[str, Any]) -> Dict[str, Any]:
"""Reads in image frames and returns the density map and crowd count.
Args:
inputs (dict): Dictionary of inputs with key "img".
Returns:
outputs (dict): csrnet output in dictionary format with keys
"density_map" and "count".
"""
image = cv2.cvtColor(inputs["img"], cv2.COLOR_BGR2RGB)
density_map, crowd_count = self.model.predict(image)
outputs = {"density_map": density_map, "count": crowd_count}
return outputs
def _get_config_types(self) -> Dict[str, Any]:
"""Returns dictionary mapping the node's config keys to respective types."""
return {"model_type": str, "weights_parent_dir": Optional[str], "width": int}