def testCase1(self):
start_disp = -4
dilation = 2
alpha = 1.0
normalize = True
max_disp = 9
h, w = 2, 2
d = (max_disp + dilation - 1) // dilation
cfg = Config(
dict(model=dict(
disp_predictor=dict(
type=self.pred_type,
# the maximum disparity of disparity search range
max_disp=max_disp,
# disparity sample radius
radius=self.radius,
# the start disparity of disparity search range
start_disp=start_disp,
# the step between near disparity sample
dilation=dilation,
# the step between near disparity sample when local sampling
radius_dilation=self.radius_dilation,
# the temperature coefficient of soft argmin
alpha=alpha,
# whether normalize the estimated cost volume
normalize=normalize,
), )))
cfg.model.update(
disp_predictor=kick_out_none_keys(cfg.model.disp_predictor))
cost = torch.ones(1, d, h, w).to(self.device)
cost.requires_grad = True
print('*' * 60)
print('Cost volume:')
print(cost)
disp_predictor = build_disp_predictor(cfg).to(self.device)
print(disp_predictor)
disp = disp_predictor(cost)
print('*' * 60)
print('Regressed disparity map :')
print(disp)
# soft argmin
if self.pred_type == 'DEFAULT':
print('*' * 60)
print('Test directly providing disparity samples')
end_disp = start_disp + max_disp - 1
# generate disparity samples
disp_samples = torch.linspace(start_disp, end_disp, d).repeat(1, h, w, 1).\
permute(0, 3, 1, 2).contiguous().to(cost.device)
disp = disp_predictor(cost, disp_samples)
print('Regressed disparity map :')
print(disp)
def __init__(self, cfg):
super(GeneralizedStereoModel, self).__init__()
self.cfg = cfg.copy()
self.max_disp = cfg.model.max_disp
self.scale = cfg.model.backbone.scale
self.backbone = build_backbone(cfg)
self.cost_processor = build_cost_processor(cfg)
# confidence measurement network
self.cmn = None
if 'cmn' in cfg.model:
self.cmn = build_cmn(cfg)
self.disp_predictor = build_disp_predictor(cfg)
self.loss_evaluator = make_gsm_loss_evaluator(cfg)