This repository is re-implementation of CSQ(Central Similarity Quantization) with MFNet(Multi-Fiber Net)
CSQ is a methodology for video retrieval, and in this repository, we use the UCF101 dataset to conduct video retrieval.
- Get the code.
git clone https://github.com/Jo-won/CSQ_pytorch.git
-
Check requirement.txt
-
Go to Wandb and create an account.
Running this code requires UCF101 dataset. (available here)
The data path can be checked below.
Please move trainlist01_ilp.txt
and testlist01_ilp.txt
inside ucf101_txt folder
<DATA_PATH>
+-- DATA_ROOT
| +-- video
| | +-- ApplyEyeMakeup
| | | +-- v_ApplyEyeMakeup_g01_c01.avi
| | | +-- ...
| | +-- ApplyLipstick
| | +-- ...
| +-- ucfTrainTestlist
| | +-- trainlist01_ilp.txt
| | +-- testlist01_ilp.txt
- Train & Vaild
Please set themodel_root
folder to be saved atopt.py
,
set thedata_root
in thetrain_ucf101_base.sh
file,
and then use the command below.
Validation proceeds with the same videos as the test.
bash train_ucf101_base.sh
- Result
When CSQ 64bits,
mAP@100 (CSQ 64bits) | Original Paper | Our re-implementation |
---|---|---|
UCF101 | 0.874 | 0.934 |
- CSQ official code : https://github.com/yuanli2333/Hadamard-Matrix-for-hashing
- MFNet official code : https://github.com/cypw/PyTorch-MFNet