GroundNLQ @ Ego4D Natural Language Queries Challenge 2023

Techical report

TL;DR: GroundNLQ won the first place at the Ego4D Natural Language Queries Challenge at CVPR23. Technically, we leverage a two-stage pre-training strategy to train egocentric feature extractors and the grounding model on video narrations, and further fine-tune the model on annotated data. In addition, we introduce a novel grounding model GroundNLQ, which employs a multi-modal multi-scale grounding module for effective video and text fusion and various temporal intervals, especially for long videos.

This repo supports data pre-processing, training and evaluation of the Ego4D-NLQ dataset.

Changelog

• [2024.01.22] We release the pretrained weights and some checkpoints.
• [2023.07.08] We release the GroundNLQ codebase.
• [2023.06.19] GroundNLQ model won 1st place@ Ego4D NLQ Challenge 2023, hosted by CVPR 2023.

Table of Contents

• Preparation
  • Install dependencies
  • Prepare offline data (e.g., features and files)
• Code Overview
• Experiments
  • Pre-Train
  • From-scratch
  • Finetune
  • Inference
  • Ensemble
• Contact
• Acknowledgement

Preparation

Install-dependencies

• Follow INSTALL.md for installing necessary dependencies and compiling the code.

Prepare-offline-data

• We leverage the extracted egocentric InterVideo and EgoVLP features and CLIP textual token features, please refer to Feature_Extraction_RM.
• Please refer to File_Precess_RM and Narration_Precess_RM.
• Download the Ego4D-NLQ and model data, including files, features, pretrained-weights, checkpoints. (https://drive.google.com/file/d/1tyrVLcs0m0svA6TZpubX_VlpjAvLU18F/view?usp=sharing) (about 10G).

Code-Overview

The structure of this code repo is heavily inspired by Detectron2. Some of the main components are

• ./libs/core: Parameter configuration module.
• ./libs/datasets: Data loader and IO module.
• ./libs/modeling: Our main model with all its building blocks.
• ./libs/utils: Utility functions for training, inference, and postprocessing.

Experiments

We adopt distributed data parallel DDP and fault-tolerant distributed training with torchrun.

Pre-training

Training can be launched by running the following command:

bash tools/pretrain_ego4d_narration.sh CONFIG_FILE OUTPUT_PATH 

where CONFIG_FILE is the config file for model/dataset hyperparameter initialization, OUTPUT_PATH is the model output directory name defined by yourself.

The checkpoints and other experiment log files will be written into ckpt/OUTPUT_PATH. For more configurable options, please check our config file libs/core/config.py.

The actual command used in the experiments is

bash tools/pretrain_ego4d_narration.sh configs/ego4d_narrations_nlq_internvideo.yaml pretrain

Training-From-Scratch

Training can be launched by running the following command:

bash tools/train ego4d_finetune_head_twogpu.sh CONFIG_FILE OUTPUT_PATH CUDA_DEVICE_ID

where CUDA_DEVICE_ID is cuda device id.

The actual command used in the experiments is

bash tools/train_ego4d_finetune_head_twogpu.sh configs/ego4d_nlq_v2_internvideo_1e-4.yaml scratch_2gpu 0,1

Training-Finetune

Training can be launched by running the following command:

bash tools/train ego4d_finetune_head_onegpu.sh CONFIG_FILE RESUME_PATH OUTPUT_PATH CUDA_DEVICE_ID

where RESUME_PATH is the path of the pretrained model weights.

The actual command used in the experiments is

bash tools/train_ego4d_finetune_head_onegpu.sh configs/ego4d_nlq_v2_pretrain_finetune_internvideo_2.5e-5.yaml \ 
/s1_md0/leiji/v-zhijian/ego4d_nlq_cvpr_2023_data/pretrain_weights/epoch_005.pth.tar finetune_1gpu 0

bash tools/train_ego4d_finetune_head_onegpu.sh configs/ego4d_nlq_v2_pretrain_finetune_internvideo_2.5e-5_train+val.yaml \ 
/s1_md0/leiji/v-zhijian/ego4d_nlq_cvpr_2023_data/pretrain_weights/epoch_005.pth.tar finetune_1gpu 0

Inference

Once the model is trained, you can use the following commands for inference:

bash tools/inference_ego4d_nlq.sh CONFIG_FILE CHECKPOINT_PATH CUDA_DEVICE_ID 

where CHECKPOINT_PATH is the path to the saved checkpoint.

• The results (Recall@K at IoU = 0.3 or 0.5) on the val. set should be similar to the performance of the below table reported in the main report.
• Note that the best performance is not at the final epoch but at the middle epoch.

Metric \ Method	R@1 IoU=0.3	R@5 IoU=0.3	R@1 IoU=0.5	R@5 IoU=0.5
GroundNLQ (from scratch)	16.74	39.02	11.47	27.39
GroundNLQ (finetune)	26.98	53.56	18.83	40.00

Ensemble

We conduct post-model prediction ensemble to enhance performance for leaderboard submission. The actual command used in the experiments is

python ensemble.py

Contact

This repo is maintained by Zhijian Hou. Questions and discussions are welcome via [email protected].

Acknowledgements

This code is inspired by ActionFormer. We use the extracted egocentric InternVideo and EgoVLP features from the NaQ authors. We thank the authors for their awesome open-source contributions.