MM-DistillNet is a novel framework that is able to perform Multi-Object Detection and tracking using only ambient sound during inference time. The framework leverages on our new new MTA loss function that facilitates the distillation of information from multimodal teachers (RGB, thermal and depth) into an audio-only student network.

This repository contains the PyTorch implementation of our CVPR'2021 paper There is More than Meets the Eye: Self-Supervised Multi-Object Detection and Tracking with Sound by Distilling Multimodal Knowledge. The repository builds on PyTorch-YOLOv3 Metrics and Yet-Another-EfficientDet-Pytorch codebases.

If you find the code useful for your research, please consider citing our paper:

@article{riverahurtado2021mmdistillnet,
  title={There is More than Meets the Eye: Self-Supervised Multi-Object Detection and Tracking with Sound by Distilling Multimodal Knowledge},
  author={Rivera Valverde, Francisco and Valeria Hurtado, Juana and Valada, Abhinav},
  booktitle={Proceedings of the IEEE/CVF conference on computer vision and pattern recognition},
  year={2021}
}

http://rl.uni-freiburg.de/research/multimodal-distill

• Linux
• Python 3.7
• PyTorch 1.3
• CUDA 10.1

IMPORTANT NOTE: These requirements are not necessarily mandatory. However, we have only tested the code under the above settings and cannot provide support for other setups.

a. Create a conda virtual environment.

git clone https://github.com/robot-learning-freiburg/MM-DistillNet.git
cd MM-DistillNet
conda create -n mmdistillnet_env
conda activate mmdistillnet_env

b. Install dependencies

pip install -r requirements.txt

We also supply our large-scale multimodal dataset with over 113,000 time-synchronized frames of RGB, depth, thermal, and audio modalities, available at http://multimodal-distill.cs.uni-freiburg.de/#dataset

Please make sure the data is available in the directory under the name data.

The binary download contains the expected folder format for our scripts to work. The path where the binary was extracted must be updated in the configuration files, in this case configs/mm-distillnet.cfg.

You will also need to download our trained teacher-models available here. Kindly download this files and have them available in the current directory, with the name of trained_models. The directory structure should look something like this:

>ls
configs/  evaluate.py  images/  LICENSE  logs/  mp3_to_pkl.py  README.md  requirements.txt  setup.cfg  src/  train.py trained_models/

>ls trained_models
LICENSE.txt              README.txt                             yet-another-efficientdet-d2-embedding.pth  yet-another-efficientdet-d2-rgb.pth
mm-distillnet.0.pth.tar  yet-another-efficientdet-d2-depth.pth  yet-another-efficientdet-d2.pth            yet-another-efficientdet-d2-thermal.pth

Additionally, the file configs/mm-distillnet.cfg contains support for different parallelization strategies and GPU/CPU support (using PyTorch's DataParallel and DistributedDataParallel)

Due to disk space constraints, we provide a mp3 version of the audio files. Librosa is known to be slow with mp3 files, so we also provide a mp3->pickle conversion utility. The idea is, that before training we convert the audio files to a spectogram and store it to a pickle file.

mp3_to_pkl.py --dir <path to the dataset>

Edit the config file appropriately in configs folder. Our best recipe is found under configs/mm-distillnet.cfg.

python train.py --config <path to a config file>

To run the full dataset We our method using 4 GPUs with 2.4 Gb memory each (The expected runtime is 7 days). After training, the best model would be stored under <exp_name>/best.pth.tar. This file can be used to evaluate the performance of the model.

Evaluate the performance of the model (Our best model can be found under trained_models/mm-distillnet.0.pth.tar):

python evaluate.py --config <path to a config file> --checkpoint <path to checkpoint file to evaluate>

The evaluation results of our method, after bayesian optimization, are (more details can be found in the paper):

Our best pre-trained model can be found on the dataset installation path.

We have used utility functions from other open-source projects. We especially thank the authors of:

• yTorch-YOLOv3
• Yet Another EfficientDet

• Abhinav Valada
• Juana Valeria Hurtado
• Francisco Rivera

For academic usage, the code is released under the GPLv3 license. For any commercial purpose, please contact the authors.