Summary: C++ Implementation of algorithm to compute 3D bounding boxes for batches of bboxes of shape (N, 8, 3) and (M, 8, 3).
Reviewed By: gkioxari
Differential Revision: D30905190
fbshipit-source-id: 02e2cf025cd4fa3ff706ce5cf9b82c0fb5443f96
Summary:
I have implemented an exact solution for 3D IoU of oriented 3D boxes.
This file includes:
* box3d_overlap: which computes the exact IoU of box1 and box2
* box3d_overlap_sampling: which computes an approximate IoU of box1 and box2 by sampling points within the boxes
Note that both implementations currently do not support batching.
Our exact IoU implementation is based on the fact that the intersecting shape of the two 3D boxes will be formed by segments of the surface of the boxes. Our algorithm computes these segments by reasoning whether triangles of one box are within the second box and vice versa. We deal with intersecting triangles by clipping them.
Reviewed By: gkioxari
Differential Revision: D30667497
fbshipit-source-id: 2f747f410f90b7f854eeaf3036794bc3ac982917
Summary:
Copy some descriptions of renderer parameters to more places so they are easier to find.
Also a couple of small corrections, and make RasterizationSettings a dataclass.
Reviewed By: nikhilaravi, patricklabatut
Differential Revision: D30899822
fbshipit-source-id: 805cf366acb7d51cb308fa574deff0657c199673
Summary: Attempt to fix#659, an observation that the rasterizer is nondeterministic, by resolving tied faces by picking those with lower index.
Reviewed By: nikhilaravi, patricklabatut
Differential Revision: D30699039
fbshipit-source-id: 39ed797eb7e9ce7370ae71259ad6b757f9449923
Summary: Unlike other cu files, sigmoid_alpha_blend uses torch/extension.h. Avoid for possible build speed win and because of a reported problem #843 on windows with CUDA 11.4.
Reviewed By: nikhilaravi
Differential Revision: D31054121
fbshipit-source-id: 53a1f985a1695a044dfd2ee1a5b0adabdf280595
Summary: Rename sample_farthest_point.cpp to not match its CUDA equivalent.
Reviewed By: nikhilaravi
Differential Revision: D31006645
fbshipit-source-id: 135b511cbde320d2b3e07fc5b027971ef9210aa9
Summary: Remove use of nonstandard C++. Noticed on windows in issue https://github.com/facebookresearch/pytorch3d/issues/843. (We use `__restrict__` in CUDA, where it is fine, even on windows)
Reviewed By: nikhilaravi
Differential Revision: D31006516
fbshipit-source-id: 929ba9b3216cb70fad3ffa3274c910618d83973f
Summary:
CUDA implementation of farthest point sampling algorithm.
## Visual comparison
Compared to random sampling, farthest point sampling gives better coverage of the shape.
{F658631262}
## Reduction
Parallelized block reduction to find the max value at each iteration happens as follows:
1. First split the points into two equal sized parts (e.g. for a list with 8 values):
`[20, 27, 6, 8 | 11, 10, 2, 33]`
2. Use half of the thread (4 threads) to compare pairs of elements from each half (e.g elements [0, 4], [1, 5] etc) and store the result in the first half of the list:
`[20, 27, 6, 33 | 11, 10, 2, 33]`
Now we no longer care about the second part but again divide the first part into two
`[20, 27 | 6, 33| -, -, -, -]`
Now we can use 2 threads to compare the 4 elements
4. Finally we have gotten down to a single pair
`[20 | 33 | -, - | -, -, -, -]`
Use 1 thread to compare the remaining two elements
5. The max will now be at thread id = 0
`[33 | - | -, - | -, -, -, -]`
The reduction will give the farthest point for the selected batch index at this iteration.
Reviewed By: bottler, jcjohnson
Differential Revision: D30401803
fbshipit-source-id: 525bd5ae27c4b13b501812cfe62306bb003827d2
Summary:
This is a naive python implementation of the iterative farthest point sampling algorithm along with associated simple tests. The C++/CUDA implementations will follow in subsequent diffs.
The algorithm is used to subsample a pointcloud with better coverage of the space of the pointcloud.
The function has not been added to `__init__.py`. I will add this after the full C++/CUDA implementations.
Reviewed By: jcjohnson
Differential Revision: D30285716
fbshipit-source-id: 33f4181041fc652776406bcfd67800a6f0c3dd58
Summary: Fix issue #826. This is a correction to the joining of TexturesUV into a single scene.
Reviewed By: nikhilaravi
Differential Revision: D30767092
fbshipit-source-id: 03ba6a1d2f22e569d1b3641cd13ddbb8dcb87ec7
Summary:
* HAT_INV_SKEW_SYMMETRIC_TOL was a global variable and torch script gives an error when compiling that function. Move it to the function scope.
* torch script gives error when compiling acos_linear_extrapolation because bound is a union of tuple and float. The tuple version is kept in this diff.
Reviewed By: patricklabatut
Differential Revision: D30614916
fbshipit-source-id: 34258d200dc6a09fbf8917cac84ba8a269c00aef
Summary: In D30349234 (1b8d86a104) we introduced persistent=False to some register_buffer calls, which depend on PyTorch 1.6. We go back to the old behaviour for PyTorch 1.5.
Reviewed By: nikhilaravi
Differential Revision: D30731327
fbshipit-source-id: ab02ef98ee87440ef02479b72f4872b562ab85b5
Summary:
There has historically been a lot of duplication between the coarse rasterization logic for point clouds and meshes. This diff factors out the shared logic, so coarse rasterization of point clouds and meshes share the same core logic.
Previously the only difference between the coarse rasterization kernels for points and meshes was the logic for checking whether a {point / triangle} intersects a tile in the image. We implement a generic coarse rasterization kernel that takes a set of 2D bounding boxes rather than geometric primitives; we then implement separate kernels that compute 2D bounding boxes for points and triangles.
This change does not affect the Python API at all. It also should not change any rasterization behavior, since this diff is just a refactoring of the existing logic.
I see this diff as the first in a few pieces of rasterizer refactoring. Followup diffs should do the following:
- Add a check for bin overflow in the generic coarse rasterizer kernel: allocate a global scalar to flag bin overflow which kernel worker threads can write to in case they detect bin overflow. The C++ launcher function can then check this flag after the kernel returns and issue a warning to the user in case of overflow.
- As a slightly more involved mechanism, if bin overflow is detected then the coarse kernel can continue running in order to count how many elements fall into each bin, without actually writing out their indices to the coarse output tensor. Then the actual number of entries per bin can be used to re-allocate the output tensor and re-run the coarse rasterization kernel so that bin overflow can be automatically avoided.
- The unification of the coarse and fine rasterization kernels also allows us to insert an extra CUDA kernel prior to coarse rasterization that filters out primitives outside the view frustum. This would be helpful for rendering full scenes (e.g. Matterport data) where only a small piece of the mesh is actually visible at any one time.
Reviewed By: bottler
Differential Revision: D25710361
fbshipit-source-id: 9c9dea512cb339c42adb3c92e7733fedd586ce1b
Summary: Renaming parts of the mesh coarse rasterization and separating the bounding box calculation. All in preparation for sharing code with point rasterization.
Reviewed By: bottler
Differential Revision: D30369112
fbshipit-source-id: 3508c0b1239b355030cfa4038d5f3d6a945ebbf4
Summary: In preparation for sharing coarse rasterization between point clouds and meshes, move the functions to a new file. No code changes.
Reviewed By: bottler
Differential Revision: D30367812
fbshipit-source-id: 9e73835a26c4ac91f5c9f61ff682bc8218e36c6a
Summary: Fixes GitHub issue #751. The vectorized implementation of bilinear interpolation didn't properly handle the edge cases in the same way as the `grid_sample` method in PyTorch.
Reviewed By: bottler
Differential Revision: D30684208
fbshipit-source-id: edf241ecbd72d46b94ad340a4e601e26c83db88e
Summary: Replace master with main in hard coded paths or mentions in documentation
Reviewed By: bottler
Differential Revision: D30696097
fbshipit-source-id: d5ff67bb026d90d1543d10ab027f916e8361ca69
Summary:
As suggested in #802. By not persisting the _xy_grid buffer, we can allow (in some cases) a model with one image_size to be loaded from a saved model which was trained at a different resolution.
Also avoid persisting _frequencies in HarmonicEmbedding for similar reasons.
BC-break: This will cause load_state_dict, in strict mode, to complain if you try to load an old model with the new code.
Reviewed By: patricklabatut
Differential Revision: D30349234
fbshipit-source-id: d6061d1e51c9f79a78d61a9f732c9a5dfadbbb47
Summary:
Use PyTorch3D's new faster sample_pdf function instead of local Python implementation.
Also clarify deps for the Python implementation.
Reviewed By: gkioxari
Differential Revision: D30512109
fbshipit-source-id: 84cfdc00313fada37a6b29837de96f6a4646434f
Summary: New test that each subpackage of pytorch3d imports cleanly.
Reviewed By: patricklabatut
Differential Revision: D30001632
fbshipit-source-id: ca8dcac94491fc22f33602b3bbef481cba927094
Summary: Implement the sample_pdf function from the NeRF project as compiled operators.. The binary search (in searchsorted) is replaced with a low tech linear search, but this is not a problem for the envisaged numbers of bins.
Reviewed By: gkioxari
Differential Revision: D26312535
fbshipit-source-id: df1c3119cd63d944380ed1b2657b6ad81d743e49
Summary: Copy the sample_pdf operation from the NeRF project in to PyTorch3D, in preparation for optimizing it.
Reviewed By: gkioxari
Differential Revision: D27117930
fbshipit-source-id: 20286b007f589a4c4d53ed818c4bc5f2abd22833
Summary: Small fix for omitting this argument.
Reviewed By: nikhilaravi
Differential Revision: D29548610
fbshipit-source-id: f25032fab3faa2f09006f5fcf8628138555f2f20
Summary: Fixes to a couple of comments on points to volumes, make the mask work in round_points_to_volumes, and remove a duplicate rand calculation
Reviewed By: nikhilaravi
Differential Revision: D29522845
fbshipit-source-id: 86770ba37ef3942b909baf63fd73eed1399635b6
Summary:
Implementation of ball query from PointNet++. This function is similar to KNN (find the neighbors in p2 for all points in p1). These are the key differences:
- It will return the **first** K neighbors within a specified radius as opposed to the **closest** K neighbors.
- As all the points in p2 do not need to be considered to find the closest K, the algorithm is much faster than KNN when p2 has a large number of points.
- The neighbors are not sorted
- Due to the radius threshold it is not guaranteed that there will be K neighbors even if there are more than K points in p2.
- The padding value for `idx` is -1 instead of 0.
# Note:
- Some of the code is very similar to KNN so it could be possible to modify the KNN forward kernels to support ball query.
- Some users might want to use kNN with ball query - for this we could provide a wrapper function around the current `knn_points` which enables applying the radius threshold afterwards as an alternative. This could be called `ball_query_knn`.
Reviewed By: jcjohnson
Differential Revision: D30261362
fbshipit-source-id: 66b6a7e0114beff7164daf7eba21546ff41ec450
Summary: Fix to resolve GitHub issue #796 - the cameras were being passed in the renderer forward pass instead of at initialization. The rasterizer was correctly using the cameras passed in the `kwargs` for the projection, but the `cameras` are still part of the `kwargs` for the `get_screen_to_ndc_transform` and `get_ndc_to_screen_transform` functions which is causing issues about duplicate arguments.
Reviewed By: bottler
Differential Revision: D30175679
fbshipit-source-id: 547e88d8439456e728fa2772722df5fa0fe4584d
Summary: Restore compatibility with PyTorch 1.4 and 1.5, and a few lint fixes.
Reviewed By: patricklabatut
Differential Revision: D30048115
fbshipit-source-id: ee05efa7c625f6079fb06a3cc23be93e48df9433
Summary:
API fix for NDC/screen cameras and compatibility with PyTorch3D renderers.
With this new fix:
* Users can define cameras and `transform_points` under any coordinate system conventions. The transformation applies the camera K and RT to the input points, not regarding for PyTorch3D conventions. So this makes cameras completely independent from PyTorch3D renderer.
* Cameras can be defined either in NDC space or screen space. For existing ones, FoV cameras are in NDC space. Perspective/Orthographic can be defined in NDC or screen space.
* The interface with PyTorch3D renderers happens through `transform_points_ndc` which transforms points to the NDC space and assumes that input points are provided according to PyTorch3D conventions.
* Similarly, `transform_points_screen` transforms points to screen space and again assumes that input points are under PyTorch3D conventions.
* For Orthographic/Perspective cameras, if they are defined in screen space, the `get_ndc_camera_transform` allows points to be converted to NDC for use for the renderers.
Reviewed By: nikhilaravi
Differential Revision: D26932657
fbshipit-source-id: 1a964e3e7caa54d10c792cf39c4d527ba2fb2e79
Summary: This fixes a recently introduced circular import: the problem went unnoticed by having `pytorch3d.renderer` imported first...
Reviewed By: bottler
Differential Revision: D29686235
fbshipit-source-id: 4b9f2faecec2cc8347ee259cfc359dc9e4f67784
Summary: This changes only documentation. We want to be explicit that ray directions are not normalised (nor assumed to be normalised) but their magnitude is used.
Reviewed By: nikhilaravi
Differential Revision: D29845210
fbshipit-source-id: b81fb3da13a42ad20e8721ed5271fd4f3d8f5acb
Summary: Use PathManager for checking file existence, rather than assuming the path is a local file, in a couple of cases.
Reviewed By: patricklabatut
Differential Revision: D29734621
fbshipit-source-id: e2236a7c2c50ba6916936a4d786abd601205b519
Summary: A bad env var check meant these tests were not being run. Fix that, and fix the copyright test for the new message format.
Reviewed By: patricklabatut
Differential Revision: D29734562
fbshipit-source-id: a1a9bb68901b09c71c7b4ff81a04083febca8d50
Summary:
# Background
There is an unstable error during training (it can happen after several minutes or after several hours).
The error is connected to `torch.det()` function in `_check_valid_rotation_matrix()`.
if I remove the function `torch.det()` in `_check_valid_rotation_matrix()` or remove the whole functions `_check_valid_rotation_matrix()` the error is disappeared (D29555876).
# Solution
Replace `torch.det()` with manual implementation for 3x3 matrix.
Reviewed By: patricklabatut
Differential Revision: D29655924
fbshipit-source-id: 41bde1119274a705ab849751ece28873d2c45155
Summary:
# Make `transform3d.py` a little bit better (performance and code quality)
## 1. Add decorator `torch.no_grad()` to the function `_check_valid_rotation_matrix()`
Function `_check_valid_rotation_matrix()` is needed to identify errors during forward pass only, it's not used for gradients.
## 2. Replace two calls `to` with the single one
Reviewed By: bottler
Differential Revision: D29656501
fbshipit-source-id: 4419e24dbf436c1b60abf77bda4376fb87a593be
Summary:
Context: in the code we are releasing with CO3D dataset, we use `cuda()` on TensorProperties like Pointclouds and Cameras where we recursively move batch to a GPU. It would be good to push it to a release so we don’t need to depend on the nightly build.
Additionally, I aligned the logic of `.to("cuda")` without device index to the one of `torch.Tensor` where the current device is populated to index. It should not affect any actual use cases but some tests had to be changed.
Reviewed By: bottler
Differential Revision: D29659529
fbshipit-source-id: abe58aeaca14bacc68da3e6cf5ae07df3353e3ce
Summary: An early-return test for gradient calculation did not include the opacity gradient calculation - hence would also return early without calculating gradients even if opacity gradients are required.
Reviewed By: bottler
Differential Revision: D29505684
fbshipit-source-id: 575e820b8f58b19476b2fe3288702806733e840b
Summary: This commit adds a new camera conversion function for OpenCV style parameters to Pulsar parameters to the library. Using this function it addresses a bug reported here: https://fb.workplace.com/groups/629644647557365/posts/1079637302558095, by using the PyTorch3D->OpenCV->Pulsar chain instead of the original direct conversion function. Both conversions are well-tested and an additional test for the full chain has been added, resulting in a more reliable solution requiring less code.
Reviewed By: patricklabatut
Differential Revision: D29322106
fbshipit-source-id: 13df13c2e48f628f75d9f44f19ff7f1646fb7ebd
Summary: Use rotation matrices for OpenCV / PyTorch3D conversions: this avoids hiding issues with conversions to / from axis-angle vectors and ensure new conversion functions have a consistent interface.
Reviewed By: bottler, classner
Differential Revision: D29634099
fbshipit-source-id: 40b28357914eb563fedea60a965dcf69e848ccfa
