NeuralDiff: Segmenting 3D objects that move in egocentric videos
Project Page | Paper + Supplementary | Video
About
This repository contains the official implementation of the paper NeuralDiff: Segmenting 3D objects that move in egocentric videos by Vadim Tschernezki, Diane Larlus and Andrea Vedaldi. Published at 3DV21.
Given a raw video sequence taken from a freely-moving camera, we study the problem of decomposing the observed 3D scene into a static background and a dynamic foreground containing the objects that move in the video sequence. This task is reminiscent of the classic background subtraction problem, but is significantly harder because all parts of the scene, static and dynamic, generate a large apparent motion due to the camera large viewpoint change. In particular, we consider egocentric videos and further separate the dynamic component into objects and the actor that observes and moves them. We achieve this factorization by reconstructing the video via a triple-stream neural rendering network that explains the different motions based on corresponding inductive biases. We demonstrate that our method can successfully separate the different types of motion, outperforming recent neural rendering baselines at this task, and can accurately segment moving objects. We do so by assessing the method empirically on challenging videos from the EPIC-KITCHENS dataset which we augment with appropriate annotations to create a new benchmark for the task of dynamic object segmentation on unconstrained video sequences, for complex 3D environments.
Getting started
We provide an environment config file for anaconda. You can install and activate it with the following commands:
conda env create -f environment.yaml
conda activate neuraldiff
After that, you can initialise the repository with pretrained models and the data through sh get_started.sh (this will download and extract everything you need to train and evaluate models). After that you can proceed directly to Reproducing results. If you should have any trouble with that, then initialise the repository manually as described in the next sections (Dataset, Pretrained models).
Dataset
The EPIC-Diff dataset can be downloaded here.
After downloading, move the compressed dataset to the directory of the cloned repository (e.g. NeuralDiff). Then, apply following commands:
mkdir data
mv EPIC-Diff.tar.gz data
cd data
tar -xzvf EPIC-Diff.tar.gz
The RGB frames are hosted separately as a subset from the EPIC-Kitchens dataset. The data are available at the University of Bristol data repository, data.bris. Once downloaded, move the folders into the same directory as mentioned before (data/EPIC-Diff).
Pretrained models
We are providing model checkpoints for all 10 scenes. You can use these to
- evaluate the models with the annotations from the EPIC-Diff benchmark
- create a summary video like at the top of this README to visualise the separation of the video into background, foreground and actor
The models can be downloaded here (about 50MB in total).
Once downloaded, place ckpts.tar.gz into the main directory. Then execute tar -xzvf ckpts.tar.gz. This will create a folder ckpts with the pretrained models.
Reproducing results
Visualisations and metrics per scene
To evaluate the scene with Video ID P01_01, use the following command:
sh scripts/eval.sh rel P01_01 rel 'masks' 0 0
The results are saved in results/rel. The subfolders contain a txt file containing the mAP and PSNR scores per scene and visualisations per sample.
You can find all scene IDs in the EPIC-Diff data folder (e.g. P01_01, P03_04, … P21_01).
Average metrics over all scenes
You can calculate the average of the metrics over all scenes (Table 1 in the paper) with the following command:
sh scripts/eval.sh rel 0 0 'average' 0 0
Make sure that you have calculated the metrics per scene before proceeding with that (this command simply reads the produced metrics per scene and averages them).
Rendering a video with separation of background, foreground and actor
To visualise the different model components of a reconstructed video (as seen on top of this page) from
- the ground truth camera poses corresponding to the time of the video
- and a fixed viewpoint,
use the following command:
sh scripts/eval.sh rel P01_01 rel 'summary' 0 0
This will result in a corresponding video in the folder results/rel/P01_01/summary.
The fixed viewpoints are pre-defined and correspond to the ones that we used in the videos provided in the supplementary material. You can adjust the viewpoints in __init__.py of dataset.
Training
We provide scripts for the proposed model (including colour normalisation). To train a model for scene P01_01, use the following command.
sh scripts/train.sh P01_01
You can visualise the training with tensorboard. The logs are stored in logs.
Citation
If you found our code or paper useful, then please cite our work as follows.
@inproceedings{tschernezki21neuraldiff,
author = {Vadim Tschernezki and Diane Larlus and
Andrea Vedaldi},
booktitle = {Proceedings of the International Conference
on {3D} Vision (3DV)},
title = {{NeuralDiff}: Segmenting {3D} objects that
move in egocentric videos},
year = {2021}
}
Acknowledgements
This implementation is based on this (official NeRF) and this repository (unofficial NeRF-W).
Our dataset is based on a sub-set of frames from EPIC-Kitchens. COLMAP was used for computing 3D information for these frames and VGG Image Annotator (VIA) was used for annotating them.
