A User Interface for activity recognition and activity prediction
-
Activity recognition
- Input: RGBD image with humans in the scene rendered from SUNCG. The image shape is (H,W,4). The image values belong to [0,1]. The depth of the ceiling and outside area are 0.0, and the depth of the floor is 1.0.
- Output: Words of activity label and peak response maps corresponding to the word. The value of each pixel of the word map represents the probability of the word at that position.
-
Activity prediction
- Input: Activity label and RGBD image without human in the scene rendered from SUNCG. The image shape is (H,W,4). The image values belong to [0,1]. The depth of the ceiling and outside area are 0.0, and the depth of the floor is 1.0.
- Output: Activity map. The value of each pixel of the activity map represents the probability of the activity at that position.
- Create a fresh venv (with conda or virtualenv) and activate it:
conda create -n activity python==3.7.6
conda activate activity
# create venv use requirements.txt, then skip step 2
# conda create --name activity --file requirements.txt
- Install the requirements:
conda install -c conda-forge dash-bootstrap-components scikit-image
# conda install pytorch torchvision cudatoolkit=9.2 -c pytorch
# conda install pytorch torchvision cudatoolkit=10.1 -c pytorch
conda install pytorch torchvision cudatoolkit=10.2 -c pytorch
pip install dash-canvas==0.1.0
- Start the app:
python index.py
-
Try the app at
localhost:8050!- Activity Recognition: Drag and drop an image with humans into the upload box and wait the class buttons to show up. Click the class button to see the peak response map.
- Activity Predition: Drag and drop an image without human into the upload box and select an activity label. Then wait the preditiction result to show up.