lxn96
commented
1 year ago Thank you for your interest in our work. For FSOD, different support images have a great impact on model performance. The results presented in the paper are the average of multiple randomized experiments. You can try to generate different fewlist files for experiments. Below is my code to generate fewlist, the seed can be changed to generate different support sets.
import argparse
import random
import os
import numpy as np
import mmcv
import xml.etree.ElementTree as ET
VOC = ['aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse',
'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor']
dataset_classes = {'voc': VOC}
few_nums = [1, 2, 3, 5, 10]
def is_valid(xmlpath, cls_name):
tree = ET.parse(xmlpath)
tree_root = tree.getroot()
flag = 0
for obj in tree_root.findall('object'):
name = obj.find('name').text
if name != cls_name:
continue
flag = 1
if flag:
return True
else:
return False
def gen_cls_txt(classes, root):
print('-----------------------------------------------------------')
print('--------------- Generating class list ---------------------')
if not os.path.exists(os.path.join(root, 'ClassList')):
os.mkdir(os.path.join(root, 'ClassList'))
image_ids = mmcv.list_from_file(os.path.join(root, 'ImageSets/Main/trainval.txt'))
for class_name in classes:
print('parsing {}'.format(class_name))
list_file = os.path.join(root, 'ClassList/{}.txt'.format(class_name))
with open(list_file, 'w') as out_f:
for id in image_ids:
xml_path = os.path.join(root, 'Annotations/{}.xml'.format(id))
tree = ET.parse(xml_path)
tree_root = tree.getroot()
flag = 0
for obj in tree_root.findall('object'):
name = obj.find('name').text
if name != class_name:
continue
flag = 1
if flag:
out_f.write('{}\n'.format(id))
def gen_image_fewlist(classes, root):
gen_cls_txt(classes, root)
print('-----------------------------------------------------------')
print('----------- Generating fewlist (images) ------------------')
for i, clsname in enumerate(classes):
print('===> Processing class: {}'.format(clsname))
with open(os.path.join(root, 'ClassList/{}.txt'.format(clsname)), 'r') as f:
name_list = f.readlines()
num = max(few_nums)
random.seed(i)
# selected_list = random.sample(name_list, num)
selected_list = []
while len(selected_list) < num:
x = random.sample(name_list, num)[0]
xmlpath = os.path.join(root, 'Annotations/{}.xml'.format(x.strip()))
if not is_valid(xmlpath, clsname):
continue
selected_list.append(x)
for n in few_nums:
out_path = os.path.join(root, 'benchmark_{}shot'.format(n))
if not os.path.exists(out_path):
os.mkdir(out_path)
with open(os.path.join(out_path, '{}shot_{}_train.txt'.format(n, clsname)), 'w') as f:
for i in range(n):
f.write(selected_list[i])
def get_bbox_fewlist(root, shot, classes):
rootfile = os.path.join(root, 'ImageSets/Main/trainval.txt')
with open(rootfile) as f:
names = f.readlines()
random.seed(2018)
cls_lists = [[] for _ in range(len(classes))]
cls_counts = [0] * len(classes)
cls2id = {cls: i for i, cls in enumerate(classes)}
while min(cls_counts) < shot:
imgid = random.sample(names, 1)[0]
xmlpath = os.path.join(root, 'Annotations/{}.xml'.format(imgid.strip()))
tree = ET.parse(xmlpath)
tree_root = tree.getroot()
for obj in tree_root.findall('object'):
name = obj.find('name').text
ci = cls2id[name]
if cls_counts[ci] < shot:
cls_counts[ci] += 1
if imgid not in cls_lists[ci]:
cls_lists[ci].append(imgid)
names.remove(imgid)
return cls_lists
def gen_bbox_fewlist(classes, root):
print('-----------------------------------------------------------')
print('----------- Generating fewlist (bboxes) ------------------')
for n in few_nums:
print('===> On {} shot ...'.format(n))
filelists = get_bbox_fewlist(root, n, classes)
out_path = os.path.join(root, 'benchmark_{}shot'.format(n))
if not os.path.exists(out_path):
os.mkdir(out_path)
for i, clsname in enumerate(classes):
print(' | Processing class: {}'.format(clsname))
with open(os.path.join(out_path, 'box_{}shot_{}_train.txt'.format(n, clsname)), 'w') as f:
for name in filelists[i]:
f.write(name)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--type', type=str, default='box', choices=['box', 'img', 'both'])
parser.add_argument('--dataset', type=str, default='voc', choices=['voc'])
parser.add_argument('--dataroot', type=str, default='/workspace/data/voc')
args = parser.parse_args()
classes = dataset_classes[args.dataset]
if args.type is None or args.type == 'box':
gen_bbox_fewlist(classes, args.dataroot)
elif args.type == 'img':
gen_image_fewlist(classes, args.dataroot)
elif args.type == 'both':
gen_image_fewlist(classes, args.dataroot)
gen_bbox_fewlist(classes, args.dataroot)
Alan-lab
TFA code link: 
666wzl666
This is an excellent work, but how to reproduce the results in the paper？I follow the prompts of the readme, the environment, dataset, code and config file are all consistent, But on the voc dataset, the actual reproduced results are at least 8 points worse than the results provided by the paper，Looking forward to your reply！