mirror of
https://github.com/facebookresearch/pytorch3d.git
synced 2025-08-02 03:42:50 +08:00
5910d81b7b
Summary: Add blurpool has defined in [MIP-NeRF](https://arxiv.org/abs/2103.13415). It has been added has an option for RayPointRefiner. Reviewed By: shapovalov Differential Revision: D46356189 fbshipit-source-id: ad841bad86d2b591a68e1cb885d4f781cf26c111
93 lines
3.4 KiB
Python
93 lines
3.4 KiB
Python
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
|
# All rights reserved.
|
|
#
|
|
# This source code is licensed under the BSD-style license found in the
|
|
# LICENSE file in the root directory of this source tree.
|
|
|
|
import unittest
|
|
from itertools import product
|
|
|
|
import torch
|
|
|
|
from pytorch3d.implicitron.models.renderer.ray_point_refiner import (
|
|
apply_blurpool_on_weights,
|
|
RayPointRefiner,
|
|
)
|
|
from pytorch3d.implicitron.models.renderer.ray_sampler import ImplicitronRayBundle
|
|
from tests.common_testing import TestCaseMixin
|
|
|
|
|
|
class TestRayPointRefiner(TestCaseMixin, unittest.TestCase):
|
|
def test_simple(self):
|
|
length = 15
|
|
n_pts_per_ray = 10
|
|
|
|
for add_input_samples, use_blurpool in product([False, True], [False, True]):
|
|
ray_point_refiner = RayPointRefiner(
|
|
n_pts_per_ray=n_pts_per_ray,
|
|
random_sampling=False,
|
|
add_input_samples=add_input_samples,
|
|
blurpool_weights=use_blurpool,
|
|
)
|
|
lengths = torch.arange(length, dtype=torch.float32).expand(3, 25, length)
|
|
bundle = ImplicitronRayBundle(
|
|
lengths=lengths,
|
|
origins=None,
|
|
directions=None,
|
|
xys=None,
|
|
camera_ids=None,
|
|
camera_counts=None,
|
|
)
|
|
weights = torch.ones(3, 25, length)
|
|
refined = ray_point_refiner(bundle, weights)
|
|
|
|
self.assertIsNone(refined.directions)
|
|
self.assertIsNone(refined.origins)
|
|
self.assertIsNone(refined.xys)
|
|
expected = torch.linspace(0.5, length - 1.5, n_pts_per_ray)
|
|
expected = expected.expand(3, 25, n_pts_per_ray)
|
|
if add_input_samples:
|
|
full_expected = torch.cat((lengths, expected), dim=-1).sort()[0]
|
|
else:
|
|
full_expected = expected
|
|
self.assertClose(refined.lengths, full_expected)
|
|
|
|
ray_point_refiner_random = RayPointRefiner(
|
|
n_pts_per_ray=n_pts_per_ray,
|
|
random_sampling=True,
|
|
add_input_samples=add_input_samples,
|
|
blurpool_weights=use_blurpool,
|
|
)
|
|
refined_random = ray_point_refiner_random(bundle, weights)
|
|
lengths_random = refined_random.lengths
|
|
self.assertEqual(lengths_random.shape, full_expected.shape)
|
|
if not add_input_samples:
|
|
self.assertGreater(lengths_random.min().item(), 0.5)
|
|
self.assertLess(lengths_random.max().item(), length - 1.5)
|
|
|
|
# Check sorted
|
|
self.assertTrue(
|
|
(lengths_random[..., 1:] - lengths_random[..., :-1] > 0).all()
|
|
)
|
|
|
|
def test_apply_blurpool_on_weights(self):
|
|
weights = torch.tensor(
|
|
[
|
|
[0.5, 0.6, 0.7],
|
|
[0.5, 0.3, 0.9],
|
|
]
|
|
)
|
|
expected_weights = 0.5 * torch.tensor(
|
|
[
|
|
[0.5 + 0.6, 0.6 + 0.7, 0.7 + 0.7],
|
|
[0.5 + 0.5, 0.5 + 0.9, 0.9 + 0.9],
|
|
]
|
|
)
|
|
out_weights = apply_blurpool_on_weights(weights)
|
|
self.assertTrue(torch.allclose(out_weights, expected_weights))
|
|
|
|
def test_shapes_apply_blurpool_on_weights(self):
|
|
weights = torch.randn((5, 4, 3, 2, 1))
|
|
out_weights = apply_blurpool_on_weights(weights)
|
|
self.assertEqual((5, 4, 3, 2, 1), out_weights.shape)
|