PyTorch implementation for ACM MM2023 paper:

Joint Searching and Grounding: Multi-Granularity Video Content Retrieval

The codes are modified from TVR and MS-SL

Environments

• python 3.10
• pytorch 1.13.1
• torchvision 0.14.1
• tensorboard 2.10.0
• tqdm 4.64.1
• easydict 1.10
• h5py 3.7.0
• cuda 11.7

Data Preparation

The data split files, textual RoBERTa features and visual features of Charades-STA dataset can be downloaded from Baidu Cloud Disk or Google Drive. The directory structure is expected to be the following:

data
|-- charades
|   |-- TextData
|   |-- charades_i3d_rgb_lgi.hdf5

Training and Evaluation

Training

Run the following scripts to train JSG on the corresponding dataset.

#Add project root to PYTHONPATH (Note that you need to do this each time you start a new session.)
source setup.sh

EXP_ID=train_jsg
GPU_DEVICE_ID=0
ROOTPATH=$HOME/data

./charades.sh $EXP_ID $GPU_DEVICE_ID $ROOTPATH

Evaluation

The model is placed in the directory $ROOTPATH/$DATASET/results/$MODELDIR after training. Run the following script to evaluate it(Suppose the model is trained on Charades-STA):

DATASET=charades
EVALID=eval_jsg
ROOTPATH=$HOME/data
MODELDIR=xxx

./test.sh $DATASET $EVALID $ROOTPATH $MODELDIR

Model Performance

Video-level retrieval performance

On Charades-STA

	R@1	R@5	R@10	R@100	SumR
JSG	2.4	7.7	12.8	49.8	72.7

Event-level retrieval performance

On Charades-STA

	IoU=0.3, R@10	IoU=0.3, R@100	IoU=0.5, R@10	IoU=0.5, R@100	IoU=0.7, R@10	IoU=0.7, R@100
JSG	7.23	28.71	5.67	22.50	3.28	12.34