SVD-baselines

---

0. About the paper

This repo is the source code of the implementations of the baselines in the paper "SVD: A Large-Scale Short Video Dataset for Near-Duplicate Video Retrieval" publised on ICCV-2019. The authors are Qing-Yuan Jiang, Yi He, Gen Li, Jian Lin, Lei Li and Wu-Jun Li. If you have any questions about the source code, pls contact: linj#lamda.nju.edu.cn.

1. Running Environment

python 3
pytorch

2. Running Demos

2.0. Preliminary

Directory to store some files

path/to/data
│
└───videos/
│    │ xxx.mp4
│    │ ...
│
└───frames/
│    │ xxxx.mp4/0000.jpg
│    │ xxxx.mp4/0001.jpg
│    │ ...
│    │ xxxx.mp4/xxxx.jpg
│    │ ...
│    │ xxxy.mp4/0000.jpg
│    │ xxxy.mp4/0001.jpg
│    │ ...
│    │ xxxy.mp4/xxxx.jpg
│    │ ...
│
└───features/
│    │ frames-features.h5
│    │ videos-features.h5
│    │ ...

2.1. Preprocessing

2.1.1. Frame Extraction

Required files: videos in the folder: /path/to/data/videos/*.mp4.

Run the following command:

python videoprocess/frame_extraction.py --dataname svd

The extracted frames will be saved in the folder: /path/to/data/frames/. The total storage cost for frames is about 400G (358G on my device) when fps=1.

2.1.2. Deep Features Extraction

Required files: frames/xxx.mp4/xxxx.jpg in the folder: /path/to/data/frames.

Run the following command:

CUDA_VISIBLE_DEVICES=1 python videoprocess/deepfeatures_extraction.py --dataname svd

The extracted deep features for each video will be saved in the file: /path/to/data/features/frames-features.h5. This file is about 153G when fps=1.

2.1.3. Video Features Aggregations

Required files: frames-features.h5 in the folder: /path/to/data/features.

python videoprocess/videofeatures_extraction.py --dataname svd

The aggregated features for each will be stored in the file: /path/to/data/features/videos-features.h5. This file is about 8.8G when fps=1.

2.1.4. Evaluation for Brute Force Search.

Required files: features in the folder: /path/to/data/features/videos-features.h5.

Run the following command:

python demos/bfs_demo.py --dataname svd

The map is: 0.7537

2.2. Hashing based Method

Required file: features in the folder: /path/to/data/features/videos-features.h5.

2.2.1. LSH

Run the following command:

python demos/lsh_demo.py --dataname svd --approach lsh --bit 16

The map is: 0.0370

2.2.2. ITQ

Run the following command:

python demos/itq_demo.py --dataname svd --approach itq --bit 16

The map@16bits is: 0.0560

2.2.3. IsoH

Run the following command:

python demos/isoh_demo.py --dataname svd --approach isoh --bit 16

The map@16bits is: 0.0562

2.3. Real-Value based Method

2.3.1. CNNV

• Step 1: sampling frames for clustering:

Required files: frames-features.h5 in the folder: /path/to/data/features.

python videoprocess/cnnlv_keyframe_sampling.py --dataname svd --approach cfs

Sampled frames are stored at: /path/to/data/features/cnnlv-sampling-features.h5

• Step 2: clustering

Required files: cnnlv-sampling-features.h5 in the folder: /path/to/data/features.

python videoprocess/cnnv_clustering.py --dataname svd --approach cnnvcluster

The learned centers are stored at: /path/to/data/features/cnnv-centers.h5

• Step 3: generating cnnv features

Required files: cnnv-centers in the folder: /path/to/data/features/cnnv-centers.h5

python videoprocess/cnnv_feature_aggregation.py --dataname svd --approach cnnvfa

The learned features are stored at: /path/to/data/features/cnnv-agg-features.h5

• Step 4: evaluating cnnv

Required files: cnnv-aggregated features in the folder: /path/to/data/features/cnnv-agg-features.h5

python demos/cnnv_demo.py --dataname svd --approach cnnv

The map is: 0.1895

3. TODO list

• frame extraction
• deep feature extraction
• brute-force demo
• LSH demo
• ITQ demo
• IsoH demo
• CNNV demo