Dataset Preparation

June 24, 2025 · View on GitHub

Training data

or you can download and organize the original data from VTG-IT, TimeIT, TimePro, HTStep and LongVid.

Testing data

folder structure:

dataset                                                                           
├─ timer1          
│  ├─ annotations          
│  │  ├─ train_2k5.json                                                              
│  │  └─ tvgbench.json    
│  ├─ videos                                                                      
│  │  ├─ timerft_data                                                                 
│  │  |  ├─ xxx.mp4       
│  │  │  └─ ...
│  │  ├─ tvgbench_data                                                                      
│  │  |  ├─ xxx.mp4      
│  │  │  └─ ...
├─ activitynet                                                                    
│  ├─ annotations                                                                 
│  │  ├─ sentence_temporal_grounding                                              
│  │  │  └─ test.json                                                             
│  ├─ videos                                                                      
│  |  ├─ v_zzz_3yWpTXo.mp4       
│  │  └─ ...
├─ charades                                                                       
│  ├─ Charades_anno                                                               
│  │  └─ Charades_v1_test.csv                                                     
│  ├─ Charades_v1                                                                 
│  |  ├─ 0I0FX.mp4    
│  │  └─ ...
├─ mvbench                                       
│  ├─ json                                          
│  │  ├─ action_antonym.json                        
│  │  └─ ...               
│  ├─ videos                                       
│  │  ├─ clevrer
│  │  └─ ...
├─ tempcompass                                       
│  ├─ questions                                      
│  │  ├─ multi-choice.json                        
│  │  └─ ...               
│  ├─ videos                                       
│  │  ├─ 315784.mp4
│  │  └─ ...
├─ egoschema                                       
│  ├─ MC                                           
│  │  └─ test-00000-of-00001.parquet               
│  ├─ Subset                                       
│  │  └─ test-00000-of-00001.parquet               
│  ├─ videos                                       
│  │  ├─ 001934bb-81bd-4cd8-a574-0472ef3f6678.mp4  
│  │  └─ ...
├─ videomme                                       
│  ├─ videomme                                           
│  │  └─ test-00000-of-00001.parquet               
│  ├─ data                                       
│  │  ├─ _8lBR0E_Tx8.mp4     
└─ └─ └─ ...                                                          

ActivityNet

Download link: ActivityNet

For fine-tuning setting, before training, you need to preprocess the video data.

bash preprocess_video.sh

Specify the path to the Charades-STA dataset (video files, annotations, etc.).

Charades

Download link: Charades-v1

For fine-tuning setting, before training, you need to preprocess the video data.

bash preprocess_video.sh

Specify the path to the Charades-STA dataset (video files, annotations, etc.).

TVGBench

Download link: hf: Boshenxx/TimeR1-Dataset

MVBench

We place the annotation file of tvgbench in dataset/trainval/tvgbench.json.

Download link: hf: OpenGVLab/MVBench

VideoMME

Download link: hf: lmms-lab/MVBench

Egoschema

Download link: hf: lmms-lab/egoschema

TempCompass

Download link: hf: lmms-lab/TempCompass