File size: 5,865 Bytes
15bc41b 9564652 15bc41b 9564652 15bc41b 9564652 15bc41b |
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 |
import sys
import os
os.environ['ERPC'] = '1'
import esim_py
import torch
import cv2
import time
import pyrender
import numpy as np
import trimesh
import arg_parser
from model import TEHNetWrapper
from settings import OUTPUT_HEIGHT, OUTPUT_WIDTH, MAIN_CAMERA, REAL_TEST_DATA_PATH
def pc_normalize(pc):
pc[:, 0] /= OUTPUT_WIDTH
pc[:, 1] /= OUTPUT_HEIGHT
pc[:, :2] = 2 * pc[:, :2] - 1
ts = pc[:, 2:]
t_max = ts.max(0).values
t_min = ts.min(0).values
ts = (2 * ((ts - t_min) / (t_max - t_min))) - 1
pc[:, 2:] = ts
return pc
def process_events(events):
n_events = 2048
events[:, 2] -= events[0, 2] # normalize ts
event_grid = np.zeros((OUTPUT_HEIGHT, OUTPUT_WIDTH, 3), dtype=np.float32)
count_grid = np.zeros((OUTPUT_HEIGHT, OUTPUT_WIDTH), dtype=np.float32)
x, y, t, p = events.T
x, y = x.astype(dtype=np.int32), y.astype(dtype=np.int32)
np.add.at(event_grid, (y, x, 0), t)
np.add.at(event_grid, (y, x, 1), p == 1)
np.add.at(event_grid, (y, x, 2), p != 1)
np.add.at(count_grid, (y, x), 1)
yi, xi = np.nonzero(count_grid)
t_avg = event_grid[yi, xi, 0] / count_grid[yi, xi]
p_evn = event_grid[yi, xi, 1]
n_evn = event_grid[yi, xi, 2]
events = np.hstack([xi[:, None], yi[:, None], t_avg[:, None], p_evn[:, None], n_evn[:, None]])
sampled_indices = np.random.choice(events.shape[0], n_events)
events = events[sampled_indices]
events = torch.tensor(events, dtype=torch.float32)
coordinates = np.zeros((events.shape[0], 2))
event_frame = np.zeros((OUTPUT_HEIGHT, OUTPUT_WIDTH, 3), dtype=np.uint8)
for idx, (x, y, t_avg, p_evn, n_evn) in enumerate(events):
y, x = y.int(), x.int()
coordinates[idx] = (y, x)
event_frame[y, x, 0] = (p_evn / (p_evn + n_evn)) * 255
event_frame[y, x, -1] = (n_evn / (p_evn + n_evn)) * 255
events[:, :3] = pc_normalize(events[:, :3])
hand_data = {
'event_frame': torch.tensor(event_frame, dtype=torch.uint8),
'events': events.permute(1, 0).unsqueeze(0),
'coordinates': torch.tensor(coordinates, dtype=torch.float32)
}
return hand_data
def demo(net, device, data):
net.eval()
events = data['events']
events = events.to(device=device, dtype=torch.float32)
start_time = time.time()
with torch.no_grad():
outputs = net(events)
end_time = time.time()
N = events.shape[0]
print(end_time - start_time)
outputs['class_logits'] = outputs['class_logits'].softmax(1).argmax(1).int().cpu()
frames = list()
for idx in range(N):
hands = dict()
hands['left'] = {
'vertices': outputs['left']['vertices'][idx].cpu(),
'j3d': outputs['left']['j3d'][idx].cpu(),
}
hands['right'] = {
'vertices': outputs['right']['vertices'][idx].cpu(),
'j3d': outputs['right']['j3d'][idx].cpu(),
}
coordinates = data['coordinates']
seg_mask = np.zeros((OUTPUT_HEIGHT, OUTPUT_WIDTH, 3), dtype=np.uint8)
for edx, (y, x) in enumerate(coordinates):
y, x = y.int(), x.int()
cid = outputs['class_logits'][idx][edx]
if cid == 3:
seg_mask[y, x] = 255
else:
seg_mask[y, x, cid] = 255
hands['seg_mask'] = seg_mask
frames.append(hands)
return frames
class Ev2Hands:
def __init__(self) -> None:
arg_parser.demo()
device = torch.device('cpu')
net = TEHNetWrapper(device=device)
save_path = os.environ['CHECKPOINT_PATH']
checkpoint = torch.load(save_path, map_location=device)
net.load_state_dict(checkpoint['state_dict'], strict=True)
renderer = pyrender.OffscreenRenderer(viewport_width=OUTPUT_WIDTH, viewport_height=OUTPUT_HEIGHT)
scene = pyrender.Scene(ambient_light=(0.3, 0.3, 0.3))
light = pyrender.DirectionalLight(color=[1.0, 1.0, 1.0], intensity=0.8)
light_pose = np.eye(4)
light_pose[:3, 3] = np.array([0, -1, 1])
scene.add(light, pose=light_pose)
light_pose[:3, 3] = np.array([0, 1, 1])
scene.add(light, pose=light_pose)
light_pose[:3, 3] = np.array([1, 1, 2])
scene.add(light, pose=light_pose)
camera = MAIN_CAMERA
nc = pyrender.Node(camera=camera, matrix=np.eye(4))
scene.add_node(nc)
rot = trimesh.transformations.rotation_matrix(np.radians(180), [1, 0, 0])
mano_hands = net.hands
self.net = net
self.device = device
self.mano_hands = mano_hands
self.rot = rot
self.renderer = renderer
self.scene = scene
def __call__(self, data):
net = self.net
device = self.device
mano_hands = self.mano_hands
rot = self.rot
renderer = self.renderer
scene = self.scene
frame = demo(net=net, device=device, data=data)[0]
seg_mask = frame['seg_mask']
pred_meshes = list()
for hand_type in ['left', 'right']:
faces = mano_hands[hand_type].faces
pred_mesh = trimesh.Trimesh(frame[hand_type]['vertices'].cpu().numpy() * 1000, faces)
pred_mesh.visual.vertex_colors = [255, 0, 0]
pred_meshes.append(pred_mesh)
pred_meshes = trimesh.util.concatenate(pred_meshes)
pred_meshes.apply_transform(rot)
mesh_node = pyrender.Node(mesh=pyrender.Mesh.from_trimesh(pred_meshes))
scene.add_node(mesh_node)
pred_rgb, depth = renderer.render(scene)
scene.remove_node(mesh_node)
pred_rgb = cv2.cvtColor(pred_rgb, cv2.COLOR_RGB2BGR)
pred_rgb[pred_rgb == 255] = 0
return pred_rgb
|