TDSFT_2 / LLaVA-Next-3D /data_precessing /MGrounding_process.py
Repoaner's picture
Upload LLaVA-Next-3D/data_precessing/MGrounding_process.py with huggingface_hub
c2a0942 verified
import json
import pdb
import re
from tqdm import tqdm
data = json.load(open('/mnt/petrelfs/wangzehan/data/MGrounding-630k/Group_Grounding/gg_train_120k.json', 'r'))
llava_data = []
idx = 0
for item in tqdm(data):
conv = item['conversations']
for i in range(len(conv)//2):
question = conv[2*i]['value']
answer = conv[2*i+1]['value']
caption = question.split('<|object_ref_start|>')[-1].split('<|object_ref_end|>')[0]
if "It's in the first image" in answer:
gt = 'The object is located at: Frame-1: '
elif "It's in the second image" in answer:
gt = 'The object is located at: Frame-1: '
elif "It's in the third image" in answer:
gt = 'The object is located at: Frame-3: '
elif "It's in the fourth image" in answer:
gt = 'The object is located at: Frame-4: '
elif "It's in the fifth image" in answer:
gt = 'The object is located at: Frame-5: '
coords_str = re.findall(r'(\d+)', answer)
coords_list = [round(int(coord)/1000, 2) for coord in coords_str]
gt += str(coords_list)
llava_item = {"id": idx,
"video": item['images'],
"conversations": [{"value": f"<image> Identify the object according to the following description.\n {caption}", "from": "human"}, \
{"value": gt, "from": "gpt"}],
"metadata": {"dataset": "MGrounding_refer"}}
llava_data.append(llava_item)
idx += 1
pdb.set_trace()
json.dump(llava_data, open('extra_data/annotation/MGrounding_group_grounding_llava_format.json', 'w'))