Realcat's picture
update: major change
499e141
raw
history blame contribute delete
12.4 kB
from math import log
from loguru import logger as loguru_logger
import torch
import torch.nn.functional as F
from einops import rearrange, repeat
from kornia.utils import create_meshgrid
from src.utils.plotting import make_matching_figures
from .geometry import warp_kpts
from kornia.geometry.subpix import dsnt
from kornia.utils.grid import create_meshgrid
def static_vars(**kwargs):
def decorate(func):
for k in kwargs:
setattr(func, k, kwargs[k])
return func
return decorate
############## ↓ Coarse-Level supervision ↓ ##############
@torch.no_grad()
def mask_pts_at_padded_regions(grid_pt, mask):
"""For megadepth dataset, zero-padding exists in images"""
mask = repeat(mask, 'n h w -> n (h w) c', c=2)
grid_pt[~mask.bool()] = 0
return grid_pt
@torch.no_grad()
def spvs_coarse(data, config):
"""
Update:
data (dict): {
"conf_matrix_gt": [N, hw0, hw1],
'spv_b_ids': [M]
'spv_i_ids': [M]
'spv_j_ids': [M]
'spv_w_pt0_i': [N, hw0, 2], in original image resolution
'spv_pt1_i': [N, hw1, 2], in original image resolution
}
NOTE:
- for scannet dataset, there're 3 kinds of resolution {i, c, f}
- for megadepth dataset, there're 4 kinds of resolution {i, i_resize, c, f}
"""
# 1. misc
device = data['image0'].device
N, _, H0, W0 = data['image0'].shape
_, _, H1, W1 = data['image1'].shape
scale = config['LOFTR']['RESOLUTION'][0]
scale0 = scale * data['scale0'][:, None] if 'scale0' in data else scale
scale1 = scale * data['scale1'][:, None] if 'scale1' in data else scale
h0, w0, h1, w1 = map(lambda x: x // scale, [H0, W0, H1, W1])
# 2. warp grids
# create kpts in meshgrid and resize them to image resolution
grid_pt0_c = create_meshgrid(h0, w0, False, device).reshape(1, h0*w0, 2).repeat(N, 1, 1) # [N, hw, 2]
grid_pt0_i = scale0 * grid_pt0_c
grid_pt1_c = create_meshgrid(h1, w1, False, device).reshape(1, h1*w1, 2).repeat(N, 1, 1)
grid_pt1_i = scale1 * grid_pt1_c
# mask padded region to (0, 0), so no need to manually mask conf_matrix_gt
if 'mask0' in data:
grid_pt0_i = mask_pts_at_padded_regions(grid_pt0_i, data['mask0'])
grid_pt1_i = mask_pts_at_padded_regions(grid_pt1_i, data['mask1'])
# warp kpts bi-directionally and resize them to coarse-level resolution
# (no depth consistency check, since it leads to worse results experimentally)
# (unhandled edge case: points with 0-depth will be warped to the left-up corner)
_, w_pt0_i = warp_kpts(grid_pt0_i, data['depth0'], data['depth1'], data['T_0to1'], data['K0'], data['K1'])
_, w_pt1_i = warp_kpts(grid_pt1_i, data['depth1'], data['depth0'], data['T_1to0'], data['K1'], data['K0'])
w_pt0_c = w_pt0_i / scale1
w_pt1_c = w_pt1_i / scale0
# 3. check if mutual nearest neighbor
w_pt0_c_round = w_pt0_c[:, :, :].round()
# calculate the overlap area between warped patch and grid patch as the loss weight.
# (larger overlap area between warped patches and grid patch with higher weight)
# (overlap area range from [0, 1] rather than [0.25, 1] as the penalty of warped kpts fall on midpoint of two grid kpts)
if config.LOFTR.LOSS.COARSE_OVERLAP_WEIGHT:
w_pt0_c_error = (1.0 - 2*torch.abs(w_pt0_c - w_pt0_c_round)).prod(-1)
w_pt0_c_round = w_pt0_c_round[:, :, :].long()
nearest_index1 = w_pt0_c_round[..., 0] + w_pt0_c_round[..., 1] * w1
w_pt1_c_round = w_pt1_c[:, :, :].round().long()
nearest_index0 = w_pt1_c_round[..., 0] + w_pt1_c_round[..., 1] * w0
# corner case: out of boundary
def out_bound_mask(pt, w, h):
return (pt[..., 0] < 0) + (pt[..., 0] >= w) + (pt[..., 1] < 0) + (pt[..., 1] >= h)
nearest_index1[out_bound_mask(w_pt0_c_round, w1, h1)] = 0
nearest_index0[out_bound_mask(w_pt1_c_round, w0, h0)] = 0
loop_back = torch.stack([nearest_index0[_b][_i] for _b, _i in enumerate(nearest_index1)], dim=0)
correct_0to1 = loop_back == torch.arange(h0*w0, device=device)[None].repeat(N, 1)
correct_0to1[:, 0] = False # ignore the top-left corner
# 4. construct a gt conf_matrix
conf_matrix_gt = torch.zeros(N, h0*w0, h1*w1, device=device)
b_ids, i_ids = torch.where(correct_0to1 != 0)
j_ids = nearest_index1[b_ids, i_ids]
conf_matrix_gt[b_ids, i_ids, j_ids] = 1
data.update({'conf_matrix_gt': conf_matrix_gt})
# use overlap area as loss weight
if config.LOFTR.LOSS.COARSE_OVERLAP_WEIGHT:
conf_matrix_error_gt = w_pt0_c_error[b_ids, i_ids] # weight range: [0.0, 1.0]
data.update({'conf_matrix_error_gt': conf_matrix_error_gt})
# 5. save coarse matches(gt) for training fine level
if len(b_ids) == 0:
loguru_logger.warning(f"No groundtruth coarse match found for: {data['pair_names']}")
# this won't affect fine-level loss calculation
b_ids = torch.tensor([0], device=device)
i_ids = torch.tensor([0], device=device)
j_ids = torch.tensor([0], device=device)
data.update({
'spv_b_ids': b_ids,
'spv_i_ids': i_ids,
'spv_j_ids': j_ids
})
# 6. save intermediate results (for fast fine-level computation)
data.update({
'spv_w_pt0_i': w_pt0_i,
'spv_pt1_i': grid_pt1_i
})
def compute_supervision_coarse(data, config):
assert len(set(data['dataset_name'])) == 1, "Do not support mixed datasets training!"
data_source = data['dataset_name'][0]
if data_source.lower() in ['scannet', 'megadepth']:
spvs_coarse(data, config)
else:
raise ValueError(f'Unknown data source: {data_source}')
############## ↓ Fine-Level supervision ↓ ##############
@static_vars(counter = 0)
@torch.no_grad()
def spvs_fine(data, config, logger = None):
"""
Update:
data (dict):{
"expec_f_gt": [M, 2], used as subpixel-level gt
"conf_matrix_f_gt": [M, WW, WW], M is the number of all coarse-level gt matches
"conf_matrix_f_error_gt": [Mp], Mp is the number of all pixel-level gt matches
"m_ids_f": [Mp]
"i_ids_f": [Mp]
"j_ids_f_di": [Mp]
"j_ids_f_dj": [Mp]
}
"""
# 1. misc
pt1_i = data['spv_pt1_i']
W = config['LOFTR']['FINE_WINDOW_SIZE']
WW = W*W
scale = config['LOFTR']['RESOLUTION'][1]
device = data['image0'].device
N, _, H0, W0 = data['image0'].shape
_, _, H1, W1 = data['image1'].shape
hf0, wf0, hf1, wf1 = data['hw0_f'][0], data['hw0_f'][1], data['hw1_f'][0], data['hw1_f'][1] # h, w of fine feature
assert not config.LOFTR.ALIGN_CORNER, 'only support training with align_corner=False for now.'
# 2. get coarse prediction
b_ids, i_ids, j_ids = data['b_ids'], data['i_ids'], data['j_ids']
scalei0 = scale * data['scale0'][b_ids] if 'scale0' in data else scale
scalei1 = scale * data['scale1'][b_ids] if 'scale1' in data else scale
# 3. compute gt
m = b_ids.shape[0]
if m == 0: # special case: there is no coarse gt
conf_matrix_f_gt = torch.zeros(m, WW, WW, device=device)
data.update({'conf_matrix_f_gt': conf_matrix_f_gt})
if config.LOFTR.LOSS.FINE_OVERLAP_WEIGHT:
conf_matrix_f_error_gt = torch.zeros(1, device=device)
data.update({'conf_matrix_f_error_gt': conf_matrix_f_error_gt})
data.update({'expec_f': torch.zeros(1, 2, device=device)})
data.update({'expec_f_gt': torch.zeros(1, 2, device=device)})
else:
grid_pt0_f = create_meshgrid(hf0, wf0, False, device) - W // 2 + 0.5 # [1, hf0, wf0, 2] # use fine coordinates
grid_pt0_f = rearrange(grid_pt0_f, 'n h w c -> n c h w')
# 1. unfold(crop) all local windows
if config.LOFTR.ALIGN_CORNER is False: # even windows
assert W==8
grid_pt0_f_unfold = F.unfold(grid_pt0_f, kernel_size=(W, W), stride=W, padding=0)
grid_pt0_f_unfold = rearrange(grid_pt0_f_unfold, 'n (c ww) l -> n l ww c', ww=W**2) # [1, hc0*wc0, W*W, 2]
grid_pt0_f_unfold = repeat(grid_pt0_f_unfold[0], 'l ww c -> N l ww c', N=N)
# 2. select only the predicted matches
grid_pt0_f_unfold = grid_pt0_f_unfold[data['b_ids'], data['i_ids']] # [m, ww, 2]
grid_pt0_f_unfold = scalei0[:,None,:] * grid_pt0_f_unfold # [m, ww, 2]
# 3. warp grids and get covisible & depth_consistent mask
correct_0to1_f = torch.zeros(m, WW, device=device, dtype=torch.bool)
w_pt0_i = torch.zeros(m, WW, 2, device=device, dtype=torch.float32)
for b in range(N):
mask = b_ids == b # mask of each batch
match = int(mask.sum())
correct_0to1_f_mask, w_pt0_i_mask = warp_kpts(grid_pt0_f_unfold[mask].reshape(1,-1,2), data['depth0'][[b],...],
data['depth1'][[b],...], data['T_0to1'][[b],...],
data['K0'][[b],...], data['K1'][[b],...]) # [k, WW], [k, WW, 2]
correct_0to1_f[mask] = correct_0to1_f_mask.reshape(match, WW)
w_pt0_i[mask] = w_pt0_i_mask.reshape(match, WW, 2)
# 4. calculate the gt index of pixel-level refinement
delta_w_pt0_i = w_pt0_i - pt1_i[b_ids, j_ids][:,None,:] # [m, WW, 2]
del b_ids, i_ids, j_ids
delta_w_pt0_f = delta_w_pt0_i / scalei1[:,None,:] + W // 2 - 0.5
delta_w_pt0_f_round = delta_w_pt0_f[:, :, :].round()
if config.LOFTR.LOSS.FINE_OVERLAP_WEIGHT:
# calculate the overlap area between warped patch and grid patch as the loss weight.
w_pt0_f_error = (1.0 - 2*torch.abs(delta_w_pt0_f - delta_w_pt0_f_round)).prod(-1) # [0, 1]
delta_w_pt0_f_round = delta_w_pt0_f_round.long()
nearest_index1 = delta_w_pt0_f_round[..., 0] + delta_w_pt0_f_round[..., 1] * W # [m, WW]
# corner case: out of fine windows
def out_bound_mask(pt, w, h):
return (pt[..., 0] < 0) + (pt[..., 0] >= w) + (pt[..., 1] < 0) + (pt[..., 1] >= h)
ob_mask = out_bound_mask(delta_w_pt0_f_round, W, W)
nearest_index1[ob_mask] = 0
correct_0to1_f[ob_mask] = 0
m_ids, i_ids = torch.where(correct_0to1_f != 0)
j_ids = nearest_index1[m_ids, i_ids] # i_ids, j_ids range from [0, WW-1]
j_ids_di, j_ids_dj = j_ids // W, j_ids % W # further get the (i, j) index in fine windows of image1 (right image); j_ids_di, j_ids_dj range from [0, W-1]
m_ids, i_ids, j_ids_di, j_ids_dj = m_ids.to(torch.long), i_ids.to(torch.long), j_ids_di.to(torch.long), j_ids_dj.to(torch.long)
# expec_f_gt will be used as the gt of subpixel-level refinement
expec_f_gt = delta_w_pt0_f - delta_w_pt0_f_round
if m_ids.numel() == 0: # special case: there is no pixel-level gt
loguru_logger.warning(f"No groundtruth fine match found for local regress: {data['pair_names']}")
# this won't affect fine-level loss calculation
data.update({'expec_f': torch.zeros(1, 2, device=device)})
data.update({'expec_f_gt': torch.zeros(1, 2, device=device)})
else:
expec_f_gt = expec_f_gt[m_ids, i_ids]
data.update({"expec_f_gt": expec_f_gt})
data.update({"m_ids_f": m_ids,
"i_ids_f": i_ids,
"j_ids_f_di": j_ids_di,
"j_ids_f_dj": j_ids_dj
})
# 5. construct a pixel-level gt conf_matrix
conf_matrix_f_gt = torch.zeros(m, WW, WW, device=device, dtype=torch.bool)
conf_matrix_f_gt[m_ids, i_ids, j_ids] = 1
data.update({'conf_matrix_f_gt': conf_matrix_f_gt})
if config.LOFTR.LOSS.FINE_OVERLAP_WEIGHT:
# calculate the overlap area between warped pixel and grid pixel as the loss weight.
w_pt0_f_error = w_pt0_f_error[m_ids, i_ids]
data.update({'conf_matrix_f_error_gt': w_pt0_f_error})
if conf_matrix_f_gt.sum() == 0:
loguru_logger.info(f'no fine matches to supervise')
def compute_supervision_fine(data, config, logger=None):
data_source = data['dataset_name'][0]
if data_source.lower() in ['scannet', 'megadepth']:
spvs_fine(data, config, logger)
else:
raise NotImplementedError