README.md

---
license: cc-by-nc-sa-4.0
task_categories:
- video-text-to-text
language:
- en
tags:
- text-generaration
- video-captioning
- video-grounding
size_categories:
- 1K<n<10K
configs:
  # ===== processed data =====
  - config_name: data_processed
    data_files:
      - split: train
        path:
          - iGround_train_set_processed.jsonl
      - split: val
        path:
          - iGround_val_set_processed.jsonl
      - split: test
        path:
          - iGround_test_set_processed.jsonl

  # ===== raw data =====
  - config_name: data_raw
    data_files:
      - split: train
        path:
          - iGround_train_set_raw.jsonl
      - split: val
        path:
          - iGround_val_set_raw.jsonl
      - split: test
        path:
          - iGround_test_set_raw.jsonl

  # ===== keys =====
  - config_name: keys
    data_files:
      - split: train
        path:
          - iGround_train_set_keys.jsonl
      - split: val
        path:
          - iGround_val_set_keys.jsonl
---

This repo contains the manually annotated dataset, **iGround**, introduced in the paper *"Large-scale Pre-training for Grounded Video Caption Generation"*.

## 📦 Loading the Dataset

 You can load each configuration and split directly with the 🤗 Datasets library:

 ```python
from datasets import load_dataset

 repo = "ekazakos/iGround"

 # Available configs:
# - data_processed
# - data_raw
# - keys
#
# Each config includes the standard splits: train, val, and test.

 # data_processed: annotations after processing used to train GROVE.
#   Processing merges multiple instances of the same object type in a clip
#   into a single annotation by taking the union of all boxes for that instance.
ds_proc_train = load_dataset(repo, "data_processed", split="train")
ds_proc_val = load_dataset(repo, "data_processed", split="val")
ds_proc_test = load_dataset(repo, "data_processed", split="test")

 # data_raw: raw annotations without any processing.
#   The same object type can appear multiple times in a video,
#   with distinct bounding boxes per instance and per frame.
ds_raw_train = load_dataset(repo, "data_raw", split="train")
ds_raw_val = load_dataset(repo, "data_raw", split="val")
ds_raw_test = load_dataset(repo, "data_raw", split="test")

 # keys: contains the corresponding video_ids for the above splits.
ds_keys_train = load_dataset(repo, "keys", split="train")
ds_keys_val = load_dataset(repo, "keys", split="val")
```

## 🎥 Download iGround videos

- Fill in [this form](https://docs.google.com/forms/d/e/1FAIpQLSeqYqoFludNXy-2iQHnmIJhN8FfB0ieqnz8GWWF6hqfAEZexw/viewform?usp=header) to obtain links to the iGround videos
- Run the following script, found [here](https://github.com/ekazakos/grove/blob/main/scripts/download_iGround.sh), to download the iGround videos using the provided links
  ```bash
  bash scripts/download_iGround.sh iGround_links.txt /path/to/iground_videos_dir
  ```
- **Caution**: the links expire in 7 days

 If you use this dataset, please cite:

```bibtex
@inproceedings{kazakos2025grove,
  title     = {Large-scale Pre-training for Grounded Video Caption Generation},
  author    = {Evangelos Kazakos and Cordelia Schmid and Josef Sivic},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year      = {2025}
}
```