---
license: cc-by-nc-4.0
---

# PC2-NoiseofWeb

This repo releases data introduced in our paper

> ***PC2: Pseudo-Classification Based Pseudo-Captioning for Noisy Correspondence Learning in Cross-Modal Retrieval***  
> ***Authors**: Yue Duan, Zhangxuan Gu, Zhenzhe Ying, Lei Qi, Changhua Meng and Yinghuan Shi*
 
Quick links: [[arXiv (coming soon)]() | [Published paper (coming soon)]() | [Poster (coming soon)]() | [Zhihu (coming soon)]() | [Code download]() | [Dataset download](https://drive.google.com/file/d/1MsR9GmRDUj4NoeL4xL8TXpes51JnpsrZ/view?usp=drive_link)]

## Data Collection
We develop a new dataset named **Noise of Web (NoW)** for NCL. It contains **100K website image-meta description pairs** (**98,000 pairs for training, 1,000 for validation, and 1,000 for testing**), which are open-sourced and can be crawled by anyone. NoW has two main characteristics: *without human annotations and the noisy pairs are naturally captured*. The source data of NoW is obtained by taking screenshots when accessing web pages on mobile devices (resolution: 720 $\times$ 1280) and parsing meta descriptions in html source code. In [NCR](https://github.com/XLearning-SCU/2021-NeurIPS-NCR) (predecessor of NCL), each image in all datasets were preprocessed using Faster-RCNN detector provided by [Bottom-up Attention Model](https://github.com/peteanderson80/bottom-up-attention) to generate 36 region proposals, and each proposal was encoded as a 2048-dimensional feature. Thus, following NCR, we release our the features instead of raw images for fair comparison. However, we can not just use detection methods like Faster-RCNN to extract image features since it is trained on real-world animals and objects on MS-COCO. To tackle this, we adapt [APT](https://openaccess.thecvf.com/content/CVPR2023/papers/Gu_Mobile_User_Interface_Element_Detection_via_Adaptively_Prompt_Tuning_CVPR_2023_paper.pdf) as the detection model since it is trained on the mobile user interface data. Similar to existing datasets, we capture top 36 objects with their features for one image, that is, we can encode one image into a 36 $\times$ 768 matrix. We do not artificially control the noise ratio, as all data is obtained automatically and randomly over the web. **The estimated noise ratio of this dataset is nearly 70%**. Due to the automated and non-human curated data collection process, the noise in NoW is highly authentic and intrinsic.  

## Data Structure

```

|-- h5100k_precomp
|   |-- dev_caps_bpe.txt
|   |-- dev_caps_bert.txt
|   |-- dev_ids.txt
|   |-- dev_ims.npy
|   |-- test_caps_bpe.txt
|   |-- test_caps_bert.txt
|   |-- test_ids.txt
|   |-- test_ims.npy
|   |-- train_caps_bpe.txt
|   |-- train_caps_bert.txt
|   |-- train_ids.txt
|   |-- train_ims.npy
|-- vocab
|   |-- now100k_precomp_vocab_bert.json
|   |-- now100k_precomp_vocab_bpe.json
|   |-- now100k_precomp_vocab_jieba.json

```

Please note that since our raw data contains some sensitive business data, we only provide the **encoded image features** (\*_ims.npy) and the **token ids of the text tokenized**. For tokenizer, we use both [Tokenizers](https://github.com/huggingface/tokenizers) with [BPE](https://huggingface.co/docs/tokenizers/api/models#tokenizers.models.BPE) to produce \*_caps_bpe.txt and [BertTokenizer](https://huggingface.co/transformers/v3.0.2/model_doc/bert.html#berttokenizer) with [bert-base-multilingual-cased](https://huggingface.co/google-bert/bert-base-multilingual-cased) pre-trained model to produce \*_caps_bert.txt. **Our vocabulary size of BPE tokenizer is 10,000 and that of BertTokenizer is 32702**. \*_ids.txt records the serial number of the data in the original 500k dataset. In the future, we may process and make the original dataset public.