Learning State-Aware Visual Representations from Audible Interactions (NeurIPS 2022)

Code release for the paper Self-Supervised Representation Learning from Videos of Audible Interactions. This repo contains the PyTorch implementation and pre-trained models.

Authors: Himangi Mittal, Pedro Morgado, Unnat Jain, Abhinav Gupta

[Arxiv (Paper + Supplementary)]

Introduction

We propose a self-supervised algorithm to learn representations from egocentric video data. Recently, significant efforts have been made to capture humans interacting with their own environments as they go about their daily activities. In result, several large egocentric datasets of interaction-rich multi-modal data have emerged. However, learning representations from videos can be challenging. First, given the uncurated nature of long-form continuous videos, learning effective representations require focusing on moments in time when interactions take place. Second, visual representations of daily activities should be sensitive to changes in the state of the environment. However, current successful multi-modal learning frameworks encourage representation invariance over time. To address these challenges, we leverage audio signals to identify moments of likely interactions which are conducive to better learning. We also propose a novel self-supervised objective that learns from audible state changes caused by interactions. We validate these contributions extensively on two large-scale egocentric datasets, EPIC-Kitchens-100 and the recently released Ego4D, and show improvements on several downstream tasks, including action recognition, long-term action anticipation, and object state change classification.

For more details, please refer to our paper.

Citation

If you find our work useful in your research, please cite:

@article{mittal2022learning,
  title={Learning State-Aware Visual Representations from Audible Interactions},
  author={Mittal, Himangi and Morgado, Pedro and Jain, Unnat and Gupta, Abhinav},
  journal={arXiv preprint arXiv:2209.13583},
  year={2022}
}

Installation

(a). Clone the repository

git clone https://github.com/HimangiM/RepLAI.git

(b). Install dependencies by setting up conda

conda env create -f environment.yml

Self-Supervised Training

Training on EPIC-KITCHENS-100

Run the following command. Change the logging.name to create new log directories, logs/dataset root path by changing environment.data_dir, and dataset path by changing data.args.base_path.

sh commands/command_train_epic_kitchens.sh 

Training on Ego4D

python main.py config-name default_ssrl -m logging.name train_replai data ego4D_10s_audiopeak backbone/video avid_r2plus1d_18 backbone/audio avid_spec_cnn_9 optim.batch_size 512 environment.slurm False environment.world_size -1 environment.multiprocessing_distributed False environment.distributed True environment.dist_url env:// environment.rank -1 environment.ngpu 8 environment.workers 96 environment.data_dir ./experiments/ logging.save_freq 10 optim.epochs 100 optim.args.lr 0.05 criterion.args.clr_coeff 0.5 criterion.args.aot_coeff 0.5 data.args.base_path ego4d_data/ data.args.delta_non_overlap 0.1 optim.use_lr_scheduler True optim.lr_scheduler_args.max_lr 0.05 optim.lr_scheduler_args.total_steps 100 backbone.video.args.pretrained True backbone.audio.args.pretrained True

Evaluation

Evaluation on EPIC-KITCHENS-100

Run the following command for evaluation the downstream task of action recognition using a linear classifier. Change the logging.name to load the pre-trained model, logs/dataset root path by changing environment.data_dir, and logging.suffix to create new logs for evaluations.

sh commands/command_eval_epic_kitchens.sh 

Evaluation on Ego4D

State Change Classification and Point-of-no-return temporal localization

cd hands-and-objects/state-change-localization-classification/i3d-resnet50
python train.py --cfg configs/ssrl_keyframe_loc_release1-v2_main-experiment.yaml --extra-args MISC.OUTPUT_DIR ./log/outputs/state-change-localization-classification/run2 MISC.NUM_GPUS 1 MISC.NUM_SHARDS 8 DATA_LOADER.NUM_WORKERS 4 TRAIN.BATCH_SIZE 32 SOLVER.ACCELERATOR ddp SOLVER.BASE_LR 0.0001 MODEL.PRETRAINED /checkpoints/replai

Action recognition

cd ego4d-forecasting
tools/long_term_anticipation/ego4d_recognition.sh ${PRETRAINED_DIR}

Long-term anticipation

cd ego4d-forecasting
tools/long_term_anticipation/ego4d_forecasting.sh ${PRETRAINED_DIR}

Pre-trained models

We provide checkpoints for pre-trained models.

EPIC-KITCHENS-100

Method	Top1 Acc (Verb)	Top1 Acc (Noun)	Top5 Acc (Verb)	Top5 Acc (Noun)	Model
RepLAI w/o AStC	29.29	9.67	73.33	29.54	url
RepLAI w/o MoI	28.71	8.33	73.17	27.29	url
RepLAI (scratch)	25.75	8.12	71.25	27.29	url
RepLAI	31.71	11.25	73.54	30.54	url

Ego4D

Method	StCC: Acc	AR: Top1 Acc (Verb)	AR: Top1 Acc (Noun)	LTA: ED@(Z=20) (Verb)	LTA: ED@(Z=20) (Noun)	PNR: Err	Model
RepLAI w/o AStC	63.60	21.1	13.5	0.774	0.853	0.795	url
RepLAI w/o MoI	62.90	19.8	11.2	0.792	0.868	0.801	url
RepLAI (scratch)	66.20	22.2	14.1	0.760	0.840	0.775	url
RepLAI	66.30	22.5	14.7	0.755	0.834	0.772	url

StCC: State Change Classification, AR: Action Recognition, LTA: Long-Term Anticipation, PNR: Point-of-no-return temporal localization