dasha_f
6e1a22ba8b
assets/conveyors (274 МБ) - ленты и угловая секция NVIDIA, на которые ссылается сцена относительным путём. Раньше исключались как перекачиваемые, но без них сцена не композится из коробки. cv/ - код стереодвижков, которые вызывает control_test, без весов: * defom-stereo - рабочий бейзлайн (DEFOM vitl, вход 480, iters 24) * crestereo - второй движок, точнее по габаритам (MAE 23.5 против 32.8 мм) * fast-foundationstereo - проверялся, в бейзлайн не вошёл * circular_section.py - показатель кругового сечения, перенесён в measure_plane.py: выравнивает облако по СОБСТВЕННЫМ главным осям и режет на пяти высотах вдоль каждой. Три самодельные версии (мировые оси, одно сечение) давали хуже; результаты проверки на эталонной геометрии - в circular_section_results.json Веса по-прежнему не в репозитории - источники в MODELS.md. Наборы кадров прежних прогонов (cv/flow_*, 1.26 ГБ) исключены: это выход, а не исходники. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
213 lines
7.5 KiB
Python
213 lines
7.5 KiB
Python
import numpy as np
|
|
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
from core.utils.utils import bilinear_sampler
|
|
|
|
try:
|
|
import corr_sampler
|
|
except:
|
|
pass
|
|
|
|
try:
|
|
import alt_cuda_corr
|
|
except:
|
|
# alt_cuda_corr is not compiled
|
|
pass
|
|
|
|
class CorrSampler(torch.autograd.Function):
|
|
@staticmethod
|
|
def forward(ctx, volume, coords, radius):
|
|
ctx.save_for_backward(volume,coords)
|
|
ctx.radius = radius
|
|
corr, = corr_sampler.forward(volume, coords, radius)
|
|
return corr
|
|
@staticmethod
|
|
def backward(ctx, grad_output):
|
|
volume, coords = ctx.saved_tensors
|
|
grad_output = grad_output.contiguous()
|
|
grad_volume, = corr_sampler.backward(volume, coords, grad_output, ctx.radius)
|
|
return grad_volume, None, None
|
|
|
|
|
|
class CorrBlockFast1D:
|
|
def __init__(self, fmap1, fmap2, num_levels=4, radius=4, **kwargs):
|
|
self.num_levels = num_levels
|
|
self.radius = radius
|
|
self.corr_pyramid = []
|
|
# all pairs correlation
|
|
corr = CorrBlockFast1D.corr(fmap1, fmap2)
|
|
batch, h1, w1, dim, w2 = corr.shape
|
|
corr = corr.reshape(batch*h1*w1, dim, 1, w2)
|
|
for i in range(self.num_levels):
|
|
self.corr_pyramid.append(corr.view(batch, h1, w1, -1, w2//2**i))
|
|
corr = F.avg_pool2d(corr, [1, 2], stride=[1, 2])
|
|
|
|
def __call__(self, coords):
|
|
out_pyramid = []
|
|
bz, _, ht, wd = coords.shape
|
|
coords = coords[:, [0]]
|
|
for i in range(self.num_levels):
|
|
corr = CorrSampler.apply(self.corr_pyramid[i].squeeze(3), coords/2**i, self.radius)
|
|
out_pyramid.append(corr.view(bz, -1, ht, wd))
|
|
return torch.cat(out_pyramid, dim=1)
|
|
|
|
@staticmethod
|
|
def corr(fmap1, fmap2):
|
|
B, D, H, W1 = fmap1.shape
|
|
_, _, _, W2 = fmap2.shape
|
|
fmap1 = fmap1.view(B, D, H, W1)
|
|
fmap2 = fmap2.view(B, D, H, W2)
|
|
corr = torch.einsum('aijk,aijh->ajkh', fmap1, fmap2)
|
|
corr = corr.reshape(B, H, W1, 1, W2).contiguous()
|
|
return corr / torch.sqrt(torch.tensor(D).float())
|
|
|
|
|
|
class PytorchAlternateCorrBlock1D:
|
|
def __init__(self, fmap1, fmap2, num_levels=4, radius=4, **kwargs):
|
|
self.num_levels = num_levels
|
|
self.radius = radius
|
|
self.corr_pyramid = []
|
|
self.fmap1 = fmap1
|
|
self.fmap2 = fmap2
|
|
|
|
def corr(self, fmap1, fmap2, coords):
|
|
B, D, H, W = fmap2.shape
|
|
# map grid coordinates to [-1,1]
|
|
xgrid, ygrid = coords.split([1,1], dim=-1)
|
|
xgrid = 2*xgrid/(W-1) - 1
|
|
ygrid = 2*ygrid/(H-1) - 1
|
|
|
|
grid = torch.cat([xgrid, ygrid], dim=-1)
|
|
output_corr = []
|
|
for grid_slice in grid.unbind(3):
|
|
fmapw_mini = F.grid_sample(fmap2, grid_slice, align_corners=True)
|
|
corr = torch.sum(fmapw_mini * fmap1, dim=1)
|
|
output_corr.append(corr)
|
|
corr = torch.stack(output_corr, dim=1).permute(0,2,3,1)
|
|
|
|
return corr / torch.sqrt(torch.tensor(D).float())
|
|
|
|
def __call__(self, coords):
|
|
r = self.radius
|
|
coords = coords.permute(0, 2, 3, 1)
|
|
batch, h1, w1, _ = coords.shape
|
|
fmap1 = self.fmap1
|
|
fmap2 = self.fmap2
|
|
out_pyramid = []
|
|
for i in range(self.num_levels):
|
|
dx = torch.zeros(1)
|
|
dy = torch.linspace(-r, r, 2*r+1)
|
|
delta = torch.stack(torch.meshgrid(dy, dx), axis=-1).to(coords.device)
|
|
centroid_lvl = coords.reshape(batch, h1, w1, 1, 2).clone()
|
|
centroid_lvl[..., 0] = centroid_lvl[..., 0] / 2**i
|
|
coords_lvl = centroid_lvl + delta.view(-1, 2)
|
|
corr = self.corr(fmap1, fmap2, coords_lvl)
|
|
fmap2 = F.avg_pool2d(fmap2, [1, 2], stride=[1, 2])
|
|
out_pyramid.append(corr)
|
|
out = torch.cat(out_pyramid, dim=-1)
|
|
return out.permute(0, 3, 1, 2).contiguous().float()
|
|
|
|
|
|
class CorrBlock1D:
|
|
def __init__(self, fmap1, fmap2, coords, num_levels=4, radius=4,
|
|
scale_list=[0.25, 0.5, 2.0, 4.0], scale_corr_radius=4):
|
|
self.num_levels = num_levels
|
|
self.radius = radius
|
|
self.scale_list = scale_list
|
|
self.scale_corr_radius = scale_corr_radius
|
|
self.corr_pyramid = []
|
|
self.coords_pyramid = []
|
|
dx = torch.linspace(-radius, radius, 2*radius+1)
|
|
self.dx = dx[:, None].to(coords.device)
|
|
|
|
sdx = torch.linspace(-scale_corr_radius, scale_corr_radius, 2*scale_corr_radius+1)
|
|
self.sdx = sdx[:, None].to(coords.device)
|
|
|
|
# all pairs correlation
|
|
corr = CorrBlock1D.corr(fmap1, fmap2)
|
|
|
|
batch, h1, w1, _, w2 = corr.shape
|
|
self.batch = batch
|
|
self.h1 = h1
|
|
self.w1 = w1
|
|
self.w2 = w2
|
|
corr = corr.reshape(batch*h1*w1, 1, 1, w2)
|
|
self.coords = coords.reshape(batch*h1*w1, 1, 1, 1)
|
|
|
|
self.corr_pyramid.append(corr)
|
|
for i in range(1, self.num_levels):
|
|
corr = F.avg_pool2d(corr, [1, 2], stride=[1, 2])
|
|
self.corr_pyramid.append(corr)
|
|
|
|
def __call__(self, disp, scaling=False):
|
|
batch, _, h1, w1 = disp.shape
|
|
|
|
disp = disp.reshape(self.batch*self.h1*self.w1, 1, 1, 1)
|
|
out_pyramid = []
|
|
|
|
if scaling:
|
|
corr = self.corr_pyramid[0]
|
|
for scale in self.scale_list:
|
|
x0 = self.sdx + self.coords - scale * disp
|
|
y0 = torch.zeros_like(x0)
|
|
coords_lvl = torch.cat([x0, y0], dim=-1)
|
|
corr_s = bilinear_sampler(corr, coords_lvl)
|
|
corr_s = corr_s.view(self.batch, self.h1, self.w1, -1)
|
|
out_pyramid.append(corr_s)
|
|
else:
|
|
coords = self.coords - disp
|
|
for i in range(self.num_levels):
|
|
corr = self.corr_pyramid[i]
|
|
x0 = self.dx + coords / 2**i
|
|
y0 = torch.zeros_like(x0)
|
|
coords_lvl = torch.cat([x0, y0], dim=-1)
|
|
corr_s = bilinear_sampler(corr, coords_lvl)
|
|
corr_s = corr_s.view(self.batch, self.h1, self.w1, -1)
|
|
out_pyramid.append(corr_s)
|
|
|
|
out = torch.cat(out_pyramid, dim=-1)
|
|
return out.permute(0, 3, 1, 2).contiguous().float()
|
|
|
|
@staticmethod
|
|
def corr(fmap1, fmap2):
|
|
B, D, H, W1 = fmap1.shape
|
|
_, _, _, W2 = fmap2.shape
|
|
fmap1 = fmap1.view(B, D, H, W1)
|
|
fmap2 = fmap2.view(B, D, H, W2)
|
|
corr = torch.einsum('aijk,aijh->ajkh', fmap1, fmap2)
|
|
corr = corr.reshape(B, H, W1, 1, W2).contiguous()
|
|
return corr / torch.sqrt(torch.tensor(D).float())
|
|
|
|
|
|
class AlternateCorrBlock:
|
|
def __init__(self, fmap1, fmap2, num_levels=4, radius=4, **kwargs):
|
|
raise NotImplementedError
|
|
self.num_levels = num_levels
|
|
self.radius = radius
|
|
|
|
self.pyramid = [(fmap1, fmap2)]
|
|
for i in range(1, self.num_levels):
|
|
fmap1 = F.avg_pool2d(fmap1, 2, stride=2)
|
|
fmap2 = F.avg_pool2d(fmap2, 2, stride=2)
|
|
self.pyramid.append((fmap1, fmap2))
|
|
|
|
def __call__(self, coords):
|
|
coords = coords.permute(0, 2, 3, 1)
|
|
B, H, W, _ = coords.shape
|
|
dim = self.pyramid[0][0].shape[1]
|
|
|
|
corr_list = []
|
|
for i in range(self.num_levels):
|
|
r = self.radius
|
|
fmap1_i = self.pyramid[0][0].permute(0, 2, 3, 1).contiguous()
|
|
fmap2_i = self.pyramid[i][1].permute(0, 2, 3, 1).contiguous()
|
|
|
|
coords_i = (coords / 2**i).reshape(B, 1, H, W, 2).contiguous()
|
|
corr, = alt_cuda_corr.forward(fmap1_i, fmap2_i, coords_i, r)
|
|
corr_list.append(corr.squeeze(1))
|
|
|
|
corr = torch.stack(corr_list, dim=1)
|
|
corr = corr.reshape(B, -1, H, W)
|
|
return corr / torch.sqrt(torch.tensor(dim).float())
|