Datasets:

Languages:
English
ArXiv:
License:
ShuhuaiRen's picture
Upload 3 files
cfa73c2
raw
history blame
4.07 kB
import json
import argparse
import os
from copy import deepcopy
import pdb
import numpy as np
import random
from pathlib import Path
from collections import Counter
# read json files
def read_json(path):
with open(path, "r") as fin:
datas = json.load(fin)
annos = datas["annotations"]
return annos
def read_jsonl(path):
anno = []
with open(path, "r") as fin:
datas = fin.readlines()
for data in datas:
anno.append(json.loads(data.strip()))
return anno
def write_json(data, path):
with open(path, "w") as fout:
json.dump(data, fout)
return
def read_txt(path):
data = []
with open(path, "r") as fin:
lines = fin.readlines()
for i, line in enumerate(lines):
# e.g. AO8RW 0.0 6.9##a person is putting a book on a shelf.
line = line.strip("\n")
cap = line.split("##")[-1]
if len(cap) < 2:
continue
terms = line.split("##")[0].split(" ")
vid = terms[0] + ".mp4"
start_time = float(terms[1])
end_time = float(terms[2])
data.append({"image_id": vid, "caption": cap, "timestamp": [start_time, end_time], "id": i})
return data
def filter_sent(sent):
sent = sent.strip(" ")
if len(sent) < 2:
return False
sent = sent.replace("#", "")
return sent
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--dataset', default='qvhighlights') # anet
parser.add_argument('--anno_path', default='annotations_raw/')
parser.add_argument('--video_path', default='videos/') # ActivityNet_asr_denseCap/anet_6fps_224
parser.add_argument('--outpath', default='./')
args = parser.parse_args()
'''output data example:
{
"annotations": [
{
"image_id": "3MSZA.mp4",
"caption": "person turn a light on.",
"timestamp": [24.3, 30.4],
}],
}
'''
miss_videos = []
num_clips = []
for split in ["train", "val"]: # "val", "test"
if args.dataset == "charades":
filename = f"charades_sta_{split}.txt"
annos = read_txt(os.path.join(args.anno_path, filename))
data = {}
data["annotations"] = annos
elif args.dataset == "qvhighlights":
filename = f"highlight_{split}_release.jsonl"
annos = read_jsonl(os.path.join(args.anno_path, filename))
new_data = []
for jterm in annos:
new_term = {}
new_term["image_id"] = "v_" + jterm["vid"] + ".mp4"
# check the existance of the video
if not os.path.exists(os.path.join(args.video_path, split, new_term["image_id"])):
miss_videos.append(new_term["image_id"])
continue
new_term["id"] = jterm["qid"]
new_term["caption"] = jterm["query"]
new_term["timestamp"] = jterm["relevant_windows"]
new_term["duration"] = jterm["duration"]
new_term["relevant_clip_ids"] = jterm["relevant_clip_ids"]
new_term["saliency_scores"] = jterm["saliency_scores"]
new_data.append(new_term)
num_clips.append(int(jterm["duration"]/2))
data = {}
data["annotations"] = new_data
else:
print("Do not support this dataset!")
exit(0)
print(f"==> {args.dataset} dataset \t# examples num: {len(new_data)} \t# miss videos num: {len(miss_videos)}\t# raw data num: {len(annos)}")
out_name = "{}.caption_coco_format.json".format(split)
Path(args.outpath).mkdir(parents=True, exist_ok=True)
write_json(data, os.path.join(args.outpath, out_name))
if len(num_clips) >= 1:
count = Counter(num_clips)
# sort count dict with the clip num
print(count)
print(max(list(count.keys())))