README.md

metadata

license: mit

This is the benchmark dataset for "A Benchmark for Multi-modal Foundation Models on Low-level Vision: from Single Images to Pairs"

The structure of the jsonl files is as follows:

1. q-bench2-a1-dev.jsonl (with img_path, question, answer_candidates, correct_answer)
2. q-bench2-a1-test.jsonl (with img_path, question, answer_candidates, without correct_answer)
3. q-bench2-a2.jsonl (with img_path, empty response)

The img_path is organized as prefix + img1 + _cat_ + img2 + .jpg

For example, if the img_path is "llvisionqa_compare_dev\\00079.jpg_cat_09769.jpg.jpg", then the prefix is "llvisionqa_compare_dev", img1 is "00079.jpg", img2 is "09769.jpg".

You can use the function to get the image paths:

def get_img_names(img_path, prefix = "path_to_all_single_images"):
  img_paths = img_path.split('\\')[1][:-4].split("_cat_")
  img1_name = os.path.join(prefix,img_paths[0])
  img2_name = os.path.join(prefix,img_paths[1])
  return img1_name,img2_name

The image file structure is:

1. all_single_images: all of the single images used, Baiduyunpan download link
2. llvisionqa_compare_dev: the concatenated images for the dev subset of the perception-compare task
3. llvisionqa_compare_test: the concatenated images for the test subset of the perception-compare task
4. lldescribe_compare: the concatenated images for the description-compare task

Submission for test your own MLLM on q-bench2

1. Perception-compare task (a1): organize your jsonl file "q-bench2-a1-test_(YOUR_MLLM_NAME).jsonl" as the structure of the provided "q-bench2-a1-dev.jsonl"
2. Description-compare task (a2): simply complete the empty "response" of "q-bench2-a2.jsonl" file and rename into "q-bench2-a2_(YOUR_MLLM_NAME).jsonl"

Please contact any of the first authors to get the results of your MLLM with the submission files.

Zicheng Zhang, zzc1998@sjtu.edu.cn Haoning Wu, haoning001@e.ntu.edu.sg