Spaces:
Runtime error
Runtime error
File size: 9,079 Bytes
4d1ebf3 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 |
# import for debugging
import os
import glob
import numpy as np
from PIL import Image
# import for base_tracker
import torch
import yaml
import torch.nn.functional as F
from model.network import XMem
from inference.inference_core import InferenceCore
from util.mask_mapper import MaskMapper
from torchvision import transforms
from util.range_transform import im_normalization
import sys
sys.path.insert(0, sys.path[0]+"/../")
from tools.painter import mask_painter
from tools.base_segmenter import BaseSegmenter
from torchvision.transforms import Resize
class BaseTracker:
def __init__(self, xmem_checkpoint, device, sam_model=None, model_type=None) -> None:
"""
device: model device
xmem_checkpoint: checkpoint of XMem model
"""
# load configurations
with open("tracker/config/config.yaml", 'r') as stream:
config = yaml.safe_load(stream)
# initialise XMem
network = XMem(config, xmem_checkpoint).to(device).eval()
# initialise IncerenceCore
self.tracker = InferenceCore(network, config)
# data transformation
self.im_transform = transforms.Compose([
transforms.ToTensor(),
im_normalization,
])
self.device = device
# changable properties
self.mapper = MaskMapper()
self.initialised = False
# # SAM-based refinement
# self.sam_model = sam_model
# self.resizer = Resize([256, 256])
@torch.no_grad()
def resize_mask(self, mask):
# mask transform is applied AFTER mapper, so we need to post-process it in eval.py
h, w = mask.shape[-2:]
min_hw = min(h, w)
return F.interpolate(mask, (int(h/min_hw*self.size), int(w/min_hw*self.size)),
mode='nearest')
@torch.no_grad()
def track(self, frame, first_frame_annotation=None):
"""
Input:
frames: numpy arrays (H, W, 3)
logit: numpy array (H, W), logit
Output:
mask: numpy arrays (H, W)
logit: numpy arrays, probability map (H, W)
painted_image: numpy array (H, W, 3)
"""
if first_frame_annotation is not None: # first frame mask
# initialisation
mask, labels = self.mapper.convert_mask(first_frame_annotation)
mask = torch.Tensor(mask).to(self.device)
self.tracker.set_all_labels(list(self.mapper.remappings.values()))
else:
mask = None
labels = None
# prepare inputs
frame_tensor = self.im_transform(frame).to(self.device)
# track one frame
probs, _ = self.tracker.step(frame_tensor, mask, labels) # logits 2 (bg fg) H W
# # refine
# if first_frame_annotation is None:
# out_mask = self.sam_refinement(frame, logits[1], ti)
# convert to mask
out_mask = torch.argmax(probs, dim=0)
out_mask = (out_mask.detach().cpu().numpy()).astype(np.uint8)
num_objs = out_mask.max()
painted_image = frame
for obj in range(1, num_objs+1):
painted_image = mask_painter(painted_image, (out_mask==obj).astype('uint8'), mask_color=obj+1)
return out_mask, out_mask, painted_image
@torch.no_grad()
def sam_refinement(self, frame, logits, ti):
"""
refine segmentation results with mask prompt
"""
# convert to 1, 256, 256
self.sam_model.set_image(frame)
mode = 'mask'
logits = logits.unsqueeze(0)
logits = self.resizer(logits).cpu().numpy()
prompts = {'mask_input': logits} # 1 256 256
masks, scores, logits = self.sam_model.predict(prompts, mode, multimask=True) # masks (n, h, w), scores (n,), logits (n, 256, 256)
painted_image = mask_painter(frame, masks[np.argmax(scores)].astype('uint8'), mask_alpha=0.8)
painted_image = Image.fromarray(painted_image)
painted_image.save(f'/ssd1/gaomingqi/refine/{ti:05d}.png')
self.sam_model.reset_image()
@torch.no_grad()
def clear_memory(self):
self.tracker.clear_memory()
self.mapper.clear_labels()
if __name__ == '__main__':
# video frames (multiple objects)
video_path_list = glob.glob(os.path.join('/ssd1/gaomingqi/datasets/davis/JPEGImages/480p/horsejump-high', '*.jpg'))
video_path_list.sort()
# first frame
first_frame_path = '/ssd1/gaomingqi/datasets/davis/Annotations/480p/horsejump-high/00000.png'
# load frames
frames = []
for video_path in video_path_list:
frames.append(np.array(Image.open(video_path).convert('RGB')))
frames = np.stack(frames, 0) # N, H, W, C
# load first frame annotation
first_frame_annotation = np.array(Image.open(first_frame_path).convert('P')) # H, W, C
# ----------------------------------------------------------
# initalise tracker
# ----------------------------------------------------------
device = 'cuda:4'
XMEM_checkpoint = '/ssd1/gaomingqi/checkpoints/XMem-s012.pth'
SAM_checkpoint= '/ssd1/gaomingqi/checkpoints/sam_vit_h_4b8939.pth'
model_type = 'vit_h'
# sam_model = BaseSegmenter(SAM_checkpoint, model_type, device=device)
tracker = BaseTracker(XMEM_checkpoint, device, None, device)
# test for storage efficiency
frames = np.load('/ssd1/gaomingqi/efficiency/efficiency.npy')
first_frame_annotation = np.array(Image.open('/ssd1/gaomingqi/efficiency/template_mask.png'))
for ti, frame in enumerate(frames):
print(ti)
if ti > 200:
break
if ti == 0:
mask, prob, painted_image = tracker.track(frame, first_frame_annotation)
else:
mask, prob, painted_image = tracker.track(frame)
# save
painted_image = Image.fromarray(painted_image)
painted_image.save(f'/ssd1/gaomingqi/results/TrackA/gsw/{ti:05d}.png')
tracker.clear_memory()
for ti, frame in enumerate(frames):
print(ti)
# if ti > 200:
# break
if ti == 0:
mask, prob, painted_image = tracker.track(frame, first_frame_annotation)
else:
mask, prob, painted_image = tracker.track(frame)
# save
painted_image = Image.fromarray(painted_image)
painted_image.save(f'/ssd1/gaomingqi/results/TrackA/gsw/{ti:05d}.png')
# # track anything given in the first frame annotation
# for ti, frame in enumerate(frames):
# if ti == 0:
# mask, prob, painted_image = tracker.track(frame, first_frame_annotation)
# else:
# mask, prob, painted_image = tracker.track(frame)
# # save
# painted_image = Image.fromarray(painted_image)
# painted_image.save(f'/ssd1/gaomingqi/results/TrackA/horsejump-high/{ti:05d}.png')
# # ----------------------------------------------------------
# # another video
# # ----------------------------------------------------------
# # video frames
# video_path_list = glob.glob(os.path.join('/ssd1/gaomingqi/datasets/davis/JPEGImages/480p/camel', '*.jpg'))
# video_path_list.sort()
# # first frame
# first_frame_path = '/ssd1/gaomingqi/datasets/davis/Annotations/480p/camel/00000.png'
# # load frames
# frames = []
# for video_path in video_path_list:
# frames.append(np.array(Image.open(video_path).convert('RGB')))
# frames = np.stack(frames, 0) # N, H, W, C
# # load first frame annotation
# first_frame_annotation = np.array(Image.open(first_frame_path).convert('P')) # H, W, C
# print('first video done. clear.')
# tracker.clear_memory()
# # track anything given in the first frame annotation
# for ti, frame in enumerate(frames):
# if ti == 0:
# mask, prob, painted_image = tracker.track(frame, first_frame_annotation)
# else:
# mask, prob, painted_image = tracker.track(frame)
# # save
# painted_image = Image.fromarray(painted_image)
# painted_image.save(f'/ssd1/gaomingqi/results/TrackA/camel/{ti:05d}.png')
# # failure case test
# failure_path = '/ssd1/gaomingqi/failure'
# frames = np.load(os.path.join(failure_path, 'video_frames.npy'))
# # first_frame = np.array(Image.open(os.path.join(failure_path, 'template_frame.png')).convert('RGB'))
# first_mask = np.array(Image.open(os.path.join(failure_path, 'template_mask.png')).convert('P'))
# first_mask = np.clip(first_mask, 0, 1)
# for ti, frame in enumerate(frames):
# if ti == 0:
# mask, probs, painted_image = tracker.track(frame, first_mask)
# else:
# mask, probs, painted_image = tracker.track(frame)
# # save
# painted_image = Image.fromarray(painted_image)
# painted_image.save(f'/ssd1/gaomingqi/failure/LJ/{ti:05d}.png')
# prob = Image.fromarray((probs[1].cpu().numpy()*255).astype('uint8'))
# # prob.save(f'/ssd1/gaomingqi/failure/probs/{ti:05d}.png')
|