Bringing Online Egocentric Action Recognition into the wild

Preprint / Website / Video

Gabriele Goletto*, Mirco Planamente*, Barbara Caputo, Giuseppe Averta.

This repository contains the code for the paper: "Bringing Online Egocentric Action Recognition into the wild".

Abstract: To enable a safe and effective human-robot cooperation, it is crucial to develop models for the identification of human activities. Egocentric vision seems to be a viable solution to solve this problem, and therefore many works provide deep learning solutions to infer human actions from first person videos. However, although very promising, most of these do not consider the major challenges that comes with a realistic deployment, such as the portability of the model, the need for real-time inference, and the robustness with respect to the novel domains (i.e., new spaces, users, tasks). With this paper, we set the boundaries that egocentric vision models should consider for realistic applications, defining a novel setting of egocentric action recognition in the wild, which encourages researchers to develop novel, applications-aware solutions. We also present a new model-agnostic technique that enables the rapid repurposing of existing architectures in this new context, demonstrating the feasibility to deploy a model on a tiny device (Jetson Nano) and to perform the task directly on the edge with very low energy consumption (2.4W on average at 50 fps).

Bringing Online Egocentric Action Recognition into the wild

Getting started

Install the required libraries using the provided requirements.yml file

cd EgoWild
conda env create -f requirements.yml
conda activate egowild

Data structure

The code expects the EPIC-Kitchens data to be organized according to the following structure. You can download the EPIC-Kitchens frames using this script.

├── dataset_root
|   ├── P01_01
|   |   ├── img_0000000000.jpg
|   |   ├── x_0000000000.jpg
|   |   ├── y_0000000000.jpg
|   |   ├── .
|   |   ├── .
|   |   ├── .
|   |   ├── img_0000000100.jpg
|   |   ├── x_0000000100.jpg
|   |   ├── y_0000000100.jpg
|   ├── .
|   ├── .
|   ├── .
|   ├── P22_17
|   |   ├── img_0000000000.jpg
|   |   ├── x_0000000000.jpg
|   |   ├── y_0000000000.jpg
|   |   ├── .
|   |   ├── .
|   |   ├── .
|   |   ├── img_0000000250.jpg
|   |   ├── x_0000000250.jpg
|   |   ├── y_0000000250.jpg

Configuration

Each experiment is defined by a different configuration file (in the configs/ directory).

All the experiments inherit the configuration inserted in the field extends. Experiments may indicate additional parameters or override the default values.

Before running any experiments, the following values in the default configuration file must be updated to point to the local replica of the EPIC-Kitchens dataset.

dataset:
  RGB:
    data_path: /path/to/EPIC-KITCHENS/dataset_root
  Flow:
    data_path: /path/to/EPIC-KITCHENS/dataset_root

Python commands to launch the training for MoViNet

Standard clip-wise validation

python -m train.train \
          action=train \
          name=<model_name> \
          dataset.shift=<shift_source>-<shift_target> \
          config=movinet.yaml

Python commands to launch all the different validation proposed for MoViNet

In the following all the commands which we used for the experiments in our work will be listed.

Standard clip-wise validation

python -m train.train \
          action=validate \
          name=<model_name> \
          dataset.shift=<shift_source>-<shift_target>

Trimmed standard validation

python -m train.train \
          action=validate \
          name=<model_name> \
          dataset.shift=<shift_source>-<shift_target> \
          config=trimmed.yaml

For all boundaries detection methods in trimmed scenario just use the same arguments as in the untrimmed one (notice that ABD is not implemented for the trimmed case)

Untrimmed validation - single buffer - boundaries supervision

python -m train.train \
          action=validate \
          name=<model_name> \
          dataset.shift=<shift_source>-<shift_target> \
          config=untrimmed.yaml

Untrimmed validation - single buffer - SBL

python -m train.train \
          action=validate \
          name=<model_name> \
          dataset.shift=<shift_source>-<shift_target> \
          config=untrimmed.yaml \
          boundaries_supervision=False \
          SBL_k=<k>

Untrimmed validation - single buffer - DBL

python -m train.train \
          action=validate \
          name=<model_name> \
          dataset.shift=<shift_source>-<shift_target> \
          config=untrimmed.yaml \
          boundaries_supervision=False \
          DBL_threshold=<DBL_threshold>

Untrimmed validation - single buffer - ABD

python -m train.train \
          action=validate \
          name=<model_name> \
          dataset.shift=<shift_source>-<shift_target> \
          config=untrimmed.yaml \
          boundaries_supervision=False \
          ABD=<ABD_size>

Untrimmed validation - double buffer - boundaries supervision

python -m train.train \
          action=validate \
          name=<model_name> \
          dataset.shift=<shift_source>-<shift_target> \
          config=untrimmed_db.yaml 

Untrimmed validation - double buffer - DBL

python -m train.train \
          action=validate \
          name=<model_name> \
          dataset.shift=<shift_source>-<shift_target> \
          config=untrimmed_db.yaml \
          boundaries_supervision=False \
          DBL_threshold=<DBL_threshold>

I3D training and validation

We report all the equivalent of MoViNet config files also with I3D model. To run the equivalent train/test experiment with I3D you should just:

1. Pick the command among the MoViNet one
2. Choose the proper file in the configs folder
3. Insert the I3D config version in the command

Cite us

If you use this work, please consider citing

@article{goletto2023bringing,
  title={Bringing Online Egocentric Action Recognition into the wild},
  author={Goletto, Gabriele and Planamente, Mirco and Caputo, Barbara and Averta, Giuseppe},
  journal={IEEE Robotics and Automation Letters},
  year={2023},
  publisher={IEEE}
}

Contact 📌

If you have any question, do not hesitate to contact me:

• Gabriele Goletto: gabriele.goletto@polito.it