LLaVA-Next-3D/data_precessing/intent3d.py

# encoding: utf-8
# a = input("please input a number:")
import pdb
import torch
import json
import pdb

vigil = json.load(open('/mnt/petrelfs/wangzehan/LLaVA-Next-3D/data/processed/intent3d_id_v3_val_llava_style.json', 'r'))
# vigil = json.load(open('/mnt/petrelfs/wangzehan/LLaVA-Next-3D/data/processed/vigil3d_scannet_id_v2_val_llava.json', 'r'))
scanrefer = json.load(open('/mnt/petrelfs/wangzehan/LLaVA-Next-3D/data/processed/scanrefer_vg_val_llava_style.json', 'r'))
box_info = json.load(open('/mnt/petrelfs/wangzehan/LLaVA-Next-3D/data/metadata/scannet_val_gt_box.json', 'r'))

num_id = 0
vigil_val = []
for vigil_item in vigil:

    obj_ids = vigil_item['metadata']['object_id']
    if len(obj_ids) == 0:
        box = []
    else:
        box = [box_info[vigil_item['video']][int(obj_id)] for obj_id in obj_ids]

    desc = vigil_item['conversations'][0]['value'].split('\n')[1]
    
    new_item = {"id": num_id, "video": vigil_item['video'], "conversations": [
                {"value": f"<image>Identify the object according to the following description.\n{desc}", "from": "human"}, 
                {"value": "<ground>", "from": "gpt"}], 
                "box": box, 
                "metadata": {"dataset": "vigil", "question_type": "multiple", "ann_id": num_id, "object_id": obj_ids}}
    
    vigil_val.append(new_item)
    num_id += 1

with open('/mnt/petrelfs/wangzehan/LLaVA-Next-3D/extra_data/Other_VG/intent3d_full_vg_val_llava_style.json', 'w') as f:
    json.dump(vigil_val, f, indent=4)