Repository navigation
internal TensorRT support - #56
Conversation
9537521 to
ba7b57a
Compare
|
|
||
| pcd = o3d.geometry.PointCloud() | ||
| pcd.points = o3d.utility.Vector3dVector(inside_mask_points) | ||
| pcd.points = o3d.utility.Vector3dVector(ptcld_in_range) |
There was a problem hiding this comment.
Don't know how we missed this before, but we seemed to have forgotten to truncate point cloud points beyond depth range...
There was a problem hiding this comment.
Earlier there were two checks - one that checked that a certain percentage of the object was within range (which is what this comment points to) and then later I actually recreate the pointcloud using the max_depth
There was a problem hiding this comment.
I guess you still have that in the changes, but the refactoring of the ptcld_test meant that this needed to be switched from inside_mask_points to ptcld_in_range, like yuou did.
There was a problem hiding this comment.
That was true for the depth-generated point clouds, but I think the lidar point cloud path always lacked that step.
ba7b57a to
736ebcc
Compare
New benchmarks:
* torch16 mapped from external SSD over USB; total wall not directly comparable. |
|
Demo tested; <1.25m ATE and >=12 inliers with/without TensorRT |
| return solutions | ||
|
|
||
| def T_align(self, map1: List[Object], map2: List[Object], correspondences: np.array = None): | ||
| def T_align(self, map1: List[Object], map2: List[Object], correspondences: np.array = None, yaw_only: bool = False): |
There was a problem hiding this comment.
Gonna be a bit pedantic and rename this to xyz_yaw_only
| lc_t_std: float = 4.0 | ||
| lc_r_std: float = np.deg2rad(10.0) |
There was a problem hiding this comment.
I don't think it's a great idea mass changing a bunch of core algorithm defaults in the code. Smaller things like min_max_extent and max_depth I'm iffy on, but I think that we should keep these two defaults and epsilon the same as the paper reports.
| ### TensorRT | ||
|
|
||
| `params/demo_trt` runs the same demo with FastSAM, YOLOv8 and DINOv2 on | ||
| TensorRT instead of PyTorch: | ||
|
|
||
| ``` | ||
| pip install '.[tensorrt]' | ||
| python3 demo/demo.py -p params/demo_trt -o demo_output_trt | ||
| ``` | ||
|
|
||
| The engines are exported and compiled on first use and cached in | ||
| `$ROMAN_WEIGHTS`. | ||
|
|
There was a problem hiding this comment.
I think we make this it's own discussion outside of the demo with a link from the installation instructions
* keep core loop closure covariances and epsilon params same as paper * reference tensorrt in installation instructions, and give it its own section in the readme * rename variable in T_align from yaw_only to xyz_yaw_only
mbpeterson70
left a comment
mbpeterson70
left a comment
There was a problem hiding this comment.
@liqyn I went ahead and made some minor changes based on my comments above - if those changes look okay to you, I think it's good to be merged in!
Uh oh!
There was an error while loading. Please reload this page.