CED_Methods_Eval

To download the datasets used in this project:

Processed datasets for atomic activities (Optional):

• The IMU dataset (ESC-50 + Kitchen20): https://drive.google.com/drive/folders/1K65vnkCIdqmULeFmgJvmjzneuZey6LAq?usp=drive_link.

• The audio dataset (WISDM): https://drive.google.com/drive/folders/1WOlDvCEP5QiqrrmTEWn4F-eJHV0Drjtn?usp=sharing.

Synthetic complex event dataset:

• The complex event dataset is synthesized using the atomic activity datasets above. Here is the definition of the complex events:

• Description of the dataset:

  Consider a ($data$, $label$) sample pair from the dataset:

  • $data$ is a time series of embedding vectors, with dimension $60 \times 128$. It contains 60 timestamps of embedding vectors, and each 128-dim embedding vector encodes a 5-second window of fused IMU and audio sensor data. Hence, $data$ corresponds to a 5-minute sensor data.
  • $label$ is a time series of complex event labels, with dimension $60 \times 1$. It contains 60 timestamps of complex event labels, with 0 referring to $e_0$, 1 referring to $e_1$, 2 referring to $e_2$, and 3 referring to $e_3$. Example: if at timestamp $t$, the complex event label is $2$, then it means complex event $e_2$ is detected to "happen" in previous $0 - t$ timestamps. Note that by "happen" we refer to the exact time when the complete event pattern of complex event $e_2$ is observed.

• To use the complex event dataset, simply download training data: ce5min_train_data_{num_samples}.npy, ce5min_train_labels_{num_samples}.npy, ae2ce5min_train_data_{num_samples}.npy, ae2ce5min_train_labels_{num_samples}.npy, and test data:ce5min_test_data.npy, ce5min_test_labels.npy, ae2ce5min_test_data.npy, ae2ce5min_test_labels.npy from https://drive.google.com/drive/folders/1tATqHHPHJO9_f0fI-n4gBZtRVd1QPbX4?usp=drive_link. Here num_samplesis the number of samples synthesized in the complex event dataset, you can choose values from 4000, 6000, 8000, and 10000.