-
Notifications
You must be signed in to change notification settings - Fork 0
/
infer_bs.py
146 lines (123 loc) · 5.65 KB
/
infer_bs.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
#! /usr/bin/env python
# -*- coding: utf-8 -*-
# vim:fenc=utf-8
#
# Copyright © 2018 qiang.zhou <qiang.zhou@yz-gpu029.hogpu.cc>
#
# Distributed under terms of the MIT license.
"""
"""
from dataset.davis import DAVISDataset
from net.bfnet import BFVOSNet
import numpy as np
import torch
import argparse
from sklearn.neighbors import NearestNeighbors
from PIL import Image
import os
import cv2
from net.loss import minTripletLoss
import torch.nn as nn
from net.utils import projection_for_visualization
from net.bs import BSWrapper
from multiprocessing.dummy import Pool as ThreadPool
import time
from logger.logger import setup_logger
logger = setup_logger()
image_dims = [854, 480]
reduced_image_dims = [image_dims[1]//8+1, image_dims[0]//8+1]
embedding_vector_dims = 128
K=5
PARALLEL_BS=10
global loss_func
global bs_wrapper
mean_value = np.array([122.675, 116.669, 104.008])
def interp(x, out_size=(854, 480)):
im = Image.fromarray(x)
return np.asarray(im.resize(out_size, resample=Image.BILINEAR))
def dm(Pixels1, Pixels2):
# Pixels: feat_dim * num_pixels
sqPixels1 = np.reshape(np.sum(np.square(Pixels1), axis=0), (-1,1))
sqPixels2 = np.reshape(np.sum(np.square(Pixels2), axis=0), (-1,1))
sqDist = sqPixels1 + sqPixels2.T - 2 * np.dot(Pixels1.T, Pixels2)
sqDist = np.maximum(sqDist, 0)
return sqDist
class Postor:
def __init__(self, ref_image_embedding, ref_mask, output_dir):
self.ref_image_embedding = ref_image_embedding
self.ref_mask = ref_mask
self.output_dir = output_dir
def process(self, cur_image_embedding, cur_image, i):
ref_image_embedding = self.ref_image_embedding
ref_mask = self.ref_mask
output_dir = self.output_dir
distances = dm(cur_image_embedding, ref_image_embedding)
indices = np.argsort(distances, axis=1)[:, :K]
output_mask = np.zeros(cur_image_embedding.shape[1]).flatten()
output_mask[np.sum(ref_mask[indices], axis=1) > K/2] = 1
output_mask = output_mask.reshape(reduced_image_dims)
cv2.imwrite(os.path.join(output_dir, "{:05d}_raw.png".format(i)), output_mask * 255)
output_mask_upsample = cv2.resize(output_mask, tuple(image_dims), interpolation=cv2.INTER_LINEAR)
output_mask_final = bs_wrapper.solve(cur_image, output_mask_upsample)
cv2.imwrite(os.path.join(output_dir, "{:05d}.png".format(i)), output_mask_final * 255)
def retrieve(dataobj, vdidx, model, output_dir):
# Get reference assets
ref_data = dataobj[vdidx[0]]
ref_image_embedding = model(ref_data['image'].cuda().unsqueeze(0), ref_data['spatio_temporal_frame'].cuda().unsqueeze(0))
ref_image_embedding = ref_image_embedding.cpu().numpy().reshape((embedding_vector_dims, -1))
ref_mask = ref_data['annotation'].numpy().flatten()
postor = Postor(ref_image_embedding, ref_mask, output_dir)
# Batch inference
pointer = 1
pool = ThreadPool(processes=PARALLEL_BS)
while pointer <= len(vdidx)-1:
s = time.time()
infer_bs = min(PARALLEL_BS, len(vdidx[pointer:pointer+PARALLEL_BS]))
param_pool = []
for fid, i in enumerate(vdidx[pointer:pointer+infer_bs]):
cur_data = dataobj[i]
cur_image = cur_data['image'].cuda().unsqueeze(0)
spatio_temporal_frame = cur_data['spatio_temporal_frame'].cuda().unsqueeze(0)
cur_image_embedding = model(cur_image, spatio_temporal_frame)
cur_image_embedding = cur_image_embedding.data.cpu().numpy().reshape((embedding_vector_dims, -1))
param_pool.append([cur_image_embedding, cur_data['image'].numpy().transpose((1, 2, 0))+mean_value, pointer+fid])
pool.map(lambda x: postor.process(x[0], x[1], x[2]), param_pool)
logger.info ("Processing frame {} to {}, costing {:03f}".format(pointer, pointer+infer_bs, time.time()-s))
pointer += infer_bs
pool.close()
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--base_dir', type=str, required=True, help="Path to DAVIS directory")
parser.add_argument('--seq_name', type=str, required=True,
help='Path to directory containing input image sequence')
parser.add_argument('--model_path', type=str, required=True, help='Path to pre-trained model weight')
parser.add_argument('--output_dir', type=str, default='./output', help='Output directory to save segmentation masks')
parser.add_argument('--alpha', type=float, default=0.7)
parser.add_argument('--bsthres', type=float, default=0.3)
args = parser.parse_args()
global loss_func
global bs_wrapper
loss_func = minTripletLoss(alpha=args.alpha)
bs_wrapper = BSWrapper(imsize=image_dims, thres=args.bsthres)
dataobj = DAVISDataset(args.base_dir, image_dims, 2016, phase='val', split='val')
model = BFVOSNet(embedding_vector_dims=embedding_vector_dims)
model = model.cuda()
model.load_state_dict(torch.load(args.model_path))
model.eval()
model.freeze_bn()
def retrieve_one_seq(dataobj, seq_name, model, output_dir):
vdidx = dataobj.sequence_to_sample_idx[seq_name]
output_dir = os.path.join(output_dir, seq_name)
os.makedirs(output_dir, exist_ok=True)
with torch.no_grad():
retrieve(dataobj, vdidx, model, output_dir)
logger.info ("Start to do evaluation...")
if args.seq_name != "all":
logger.info ("Now seq {}".format(args.seq_name))
retrieve_one_seq(dataobj, args.seq_name, model, args.output_dir)
else:
for i_seq_name in dataobj.sequences:
logger.info ("Now seq {}".format(i_seq_name))
retrieve_one_seq(dataobj, i_seq_name, model, args.output_dir)
if __name__ == "__main__":
main()