# Dishcovery Mission II [CVPR 2026] — Data 来源：https://www.kaggle.com/competitions/dishcovery-mission-ii-cvpr-2026/data 采集时间：2026-10-18T14:02:12Z We provide two splits of challenge data: - Training data. - Validation data. # Suggested Training Data Training data generated by [Precision at Scale framework](https://arxiv.org/abs/2407.03463) is available on HuggingFace 🤗 [here](https://huggingface.co/collections/jesusmolrdv/cvprw-mtf-2025-vlm-challenge-data-67e3b1cd766b4836df868a1b) with two splits: - Web split. - Synthetic split. ## MTF VLM Challenge Dataset - Web Split ### Overview This dataset is part of the training data for the **CVPR Workshop Metafood VLM Challenge**. It consists of image-text pairs where the images are sourced from the web. The dataset has been carefully curated using the **Precision-at-Scale** method, ensuring high relevance and quality for domain-specific tasks. **Associated Hugging Face Repository:** `jesusmolrdv/MTF25-VLM-Challenge-Dataset-Web` ### Dataset Description * **Source:** Web-scraped images. * **Format:** The dataset hosted on Hugging Face contains a table (similar to a Pandas DataFrame) with two columns: * `url`: The URL of the original food image. * `caption`: A textual description associated with the image. * **Curation:** The selection of image-text pairs was performed using the methodology described in the "Precision at Scale" paper (see Citation section). ### How to Use / Download Since this dataset split contains URLs to images rather than the images themselves, you need to download the images using a tool like [`img2dataset`](https://github.com/rom1504/img2dataset/). 1. **Install dependencies:** ```bash pip install img2dataset datasets pyarrow pandas ``` 2. **Prepare the URL list:** First, download the dataset metadata (which contains the URLs and captions) from Hugging Face and save it as a local file (e.g., Parquet format). Run this short Python script: ```python from datasets import load_dataset import os hf_dataset_name = "jesusmolrdv/MTF25-VLM-Challenge-Dataset-Web" metadata_output_file = "mtf2025_web_metadata.parquet" # Output file for img2dataset print(f"Loading dataset metadata: {hf_dataset_name}") # Ensure you load the correct split, default is often 'train' dataset = load_dataset(hf_dataset_name, split="train") print(f"Saving metadata to {metadata_output_file}...") # Save in Parquet format, which img2dataset can read efficiently dataset.to_parquet(metadata_output_file) print("Metadata file saved successfully.") ``` This script will create a file named `mtf2025_web_metadata.parquet` in the directory where you run it. This file contains the `url` and `caption` columns needed by `img2dataset`. 3. **Download images using `img2dataset` CLI:** Now, use the `img2dataset` command in your terminal. Adjust parameters like `output_folder`, `image_size`, and `processes_count` as needed. ```bash img2dataset \ --url_list mtf2025_web_metadata.parquet \ --input_format "parquet" \ --url_col "url" \ --caption_col "caption" \ --output_format webdataset \ --output_folder ./mtf2025_web_images \ --processes_count 16 \ --thread_count 64 \ --image_size 512 \ --resize_mode keep_ratio \ --save_additional_columns '["caption"]' \ --enable_wandb False ``` > *Note: Downloading large datasets can take significant time and bandwidth. Some URLs might become inactive over time.* ### Citation If you use this dataset in your research or challenge participation, please cite the following paper describing the curation method: ``` @article{DBLP:journals/pr/RodriguezdeVeraESNR26, author={Jesús M. Rodríguez-de-Vera and Imanol G. Estepa and Ignacio Sarasúa and Bhalaji Nagarajan and Petia Radeva}, title={Precision at scale: Domain-specific datasets on-demand}, year={2026}, cdate={1767225600000}, journal={Pattern Recognit.}, volume={171}, pages={112236}, url={https://doi.org/10.1016/j.patcog.2025.112236} } ``` ## MTF VLM Challenge Dataset - Synthetic Split ### Overview This dataset is part of the training data for the **CVPR Workshop Metafood VLM Challenge**. It consists of image-text pairs where the images are synthetically generated using **[Stable Diffusion 2.1](https://huggingface.co/stabilityai/stable-diffusion-2-1)**. The dataset has been carefully curated using the **Precision-at-Scale** method, ensuring high relevance and quality for domain-specific tasks. **Associated Hugging Face Repository:** `jesusmolrdv/MTF25-VLM-Challenge-Dataset-Synth` ### Dataset Description * **Source:** Synthetic images generated via Stable Diffusion v2.1. * **Format:** The dataset hosted on Hugging Face contains a table (similar to a Pandas DataFrame or Hugging Face `Dataset` object) with two columns: * `image`: The PIL (Python Imaging Library) image object itself. * `caption`: The textual description (prompt) used to generate the image. * **Curation:** The selection/generation process was guided by the methodology described in the "Precision at Scale" paper (see Citation section). ### How to Use / Download This dataset can be directly loaded using the Hugging Face `datasets` library. The images are embedded within the dataset structure. 1. **Install Hugging Face `datasets` library:** ```bash pip install datasets Pillow ``` 2. **Load the dataset using Python:** ```python from datasets import load_dataset # Load the dataset from Hugging Face hf_dataset_name = "jesusmolrdv/MTF25-VLM-Challenge-Dataset-Synth" print(f"Loading dataset: {hf_dataset_name}...") # Load the desired split (e.g., 'train') # Use streaming=True for large datasets to avoid downloading everything at once # dataset = load_dataset(hf_dataset_name, split="train", streaming=True) dataset = load_dataset(hf_dataset_name, split="train") # Loads the entire split into memory/cache print("Dataset loaded successfully.") # Accessing data (example for non-streaming) if dataset: print(f"Number of examples in split: {len(dataset)}") # Get the first example first_example = dataset[0] img = first_example['image'] caption = first_example['caption'] print("\nFirst example:") print(f" Caption: {caption}") print(f" Image type: {type(img)}") print(f" Image mode: {img.mode}") print(f" Image size: {img.size}") # You can display the image (requires matplotlib or other library) # import matplotlib.pyplot as plt # plt.imshow(img) # plt.title(caption) # plt.axis('off') # plt.show() # Iterate through the dataset (example) # print("\nIterating through first 5 examples:") # for i, example in enumerate(dataset): # if i >= 5: # break # print(f" Example {i+1}: Caption - {example['caption'][:50]}...") # Print first 50 chars of caption # Process example['image'] and example['caption'] as needed else: print("Failed to load dataset.") # Example of iterating with streaming=True # print("\nIterating through first 5 examples (streaming):") # count = 0 # for example in dataset: # if count >= 5: # break # print(f" Example {count+1}: Caption - {example['caption'][:50]}...") # # Process example['image'] and example['caption'] # count += 1 ``` > *Note: Ensure you have enough disk space for the Hugging Face cache if not using `streaming=True`.* ### Citation If you use this dataset in your research or challenge participation, please cite the following paper describing the curation method: ``` @article{DBLP:journals/pr/RodriguezdeVeraESNR26, author={Jesús M. Rodríguez-de-Vera and Imanol G. Estepa and Ignacio Sarasúa and Bhalaji Nagarajan and Petia Radeva}, title={Precision at scale: Domain-specific datasets on-demand}, year={2026}, cdate={1767225600000}, journal={Pattern Recognit.}, volume={171}, pages={112236}, url={https://doi.org/10.1016/j.patcog.2025.112236} } ``` # Test Data - Files Provided * **Test1 Folder** - Multi-label test set. It contains all images, captions (TXT) and image order (TXT). * **Test2 Folder** - Single-label test set. It contains all images, captions (JSON) and image order (TXT). * **sample_submission.csv** - A sample submission file in the correct format.