Assume that you have a "cone" represented by normal n and cos(θ), where θ is the half angle, how do you compute the intersection of the two without involving acos?
I'm actually mostly interested in the binary classification (interesects / doesn't intersect)
I'm actually mostly interested in the binary classification (interesects / doesn't intersect)
October 2, 2025 at 1:51 AM
Assume that you have a "cone" represented by normal n and cos(θ), where θ is the half angle, how do you compute the intersection of the two without involving acos?
I'm actually mostly interested in the binary classification (interesects / doesn't intersect)
I'm actually mostly interested in the binary classification (interesects / doesn't intersect)
Prove me wrong: For all the 3R papers, there is no reason for not running a fusion over your results. The figures would be way more readable, and misalignment more obvious. The fuzziness of point-clouds hides it.
We have the technology!
2025 vs. 2016
We have the technology!
2025 vs. 2016
October 1, 2025 at 3:03 PM
Prove me wrong: For all the 3R papers, there is no reason for not running a fusion over your results. The figures would be way more readable, and misalignment more obvious. The fuzziness of point-clouds hides it.
We have the technology!
2025 vs. 2016
We have the technology!
2025 vs. 2016
Really looking forward to people starting to up the energy of their 3D reconstruction paper titles.
These guys had it figured out, i'd skip the question mark tho.
These guys had it figured out, i'd skip the question mark tho.
September 18, 2025 at 2:21 PM
Really looking forward to people starting to up the energy of their 3D reconstruction paper titles.
These guys had it figured out, i'd skip the question mark tho.
These guys had it figured out, i'd skip the question mark tho.
Not sure if useful for anybody, but I wanted a resource that lists different robust functions, their derivatives and IRLS weights in one place. So, here is a desmos workspace that does just that:
www.desmos.com/calculator/q...
www.desmos.com/calculator/q...
August 19, 2025 at 5:26 PM
Not sure if useful for anybody, but I wanted a resource that lists different robust functions, their derivatives and IRLS weights in one place. So, here is a desmos workspace that does just that:
www.desmos.com/calculator/q...
www.desmos.com/calculator/q...
Really enjoyed MP-SfM! (arxiv.org/pdf/2504.20040)
Very cool problem formulation, awesome results!
Also, a great excuse for a joke for like 3 people.
Very cool problem formulation, awesome results!
Also, a great excuse for a joke for like 3 people.
May 7, 2025 at 1:18 AM
Really enjoyed MP-SfM! (arxiv.org/pdf/2504.20040)
Very cool problem formulation, awesome results!
Also, a great excuse for a joke for like 3 people.
Very cool problem formulation, awesome results!
Also, a great excuse for a joke for like 3 people.
I keep swinging between liking python and thinking that it's not a serious programming language.
I also wouldn't underestimate my ability to write shitty code.
I also wouldn't underestimate my ability to write shitty code.
April 22, 2025 at 3:09 PM
I keep swinging between liking python and thinking that it's not a serious programming language.
I also wouldn't underestimate my ability to write shitty code.
I also wouldn't underestimate my ability to write shitty code.
Very glad to see we are using all this compute for important tasks
March 27, 2025 at 8:01 PM
Very glad to see we are using all this compute for important tasks
Excuse me, what is this black magic? May I interest you in some torch.optim.Adam, like God intended?
March 25, 2025 at 2:11 PM
Excuse me, what is this black magic? May I interest you in some torch.optim.Adam, like God intended?
Interesting, but wrong
March 16, 2025 at 2:18 PM
Interesting, but wrong
On one hand, moving to end-to-end fully learned framework will likely win at the end.
On the other hand if something like COLMAP fails, it likely is a correspondence problem, so at least you get some insight into what went wrong. Harder to reason about what went wrong in latent space.
On the other hand if something like COLMAP fails, it likely is a correspondence problem, so at least you get some insight into what went wrong. Harder to reason about what went wrong in latent space.
January 28, 2025 at 3:41 PM
On one hand, moving to end-to-end fully learned framework will likely win at the end.
On the other hand if something like COLMAP fails, it likely is a correspondence problem, so at least you get some insight into what went wrong. Harder to reason about what went wrong in latent space.
On the other hand if something like COLMAP fails, it likely is a correspondence problem, so at least you get some insight into what went wrong. Harder to reason about what went wrong in latent space.
@vincentleroy.bsky.social MASt3R retrieval module's similarity matrices seem to have odd behaviour. Self-similarity (diagonal elements) are not 1.0 and they differ between different images (slightly, mean of 0.256 and std-dev of 0.014). Any intuition why that is?
December 19, 2024 at 3:50 PM
@vincentleroy.bsky.social MASt3R retrieval module's similarity matrices seem to have odd behaviour. Self-similarity (diagonal elements) are not 1.0 and they differ between different images (slightly, mean of 0.256 and std-dev of 0.014). Any intuition why that is?
I am loving the paper! Truly great work, the formulation with the normalized rays for matching is very clever!
I am however having a bit of a hard time fully understanding the optimization for pose, and error minimized in eqn. 7. I assume this minimizes α, and not β (in the illustration below)
I am however having a bit of a hard time fully understanding the optimization for pose, and error minimized in eqn. 7. I assume this minimizes α, and not β (in the illustration below)
December 17, 2024 at 2:04 AM
I am loving the paper! Truly great work, the formulation with the normalized rays for matching is very clever!
I am however having a bit of a hard time fully understanding the optimization for pose, and error minimized in eqn. 7. I assume this minimizes α, and not β (in the illustration below)
I am however having a bit of a hard time fully understanding the optimization for pose, and error minimized in eqn. 7. I assume this minimizes α, and not β (in the illustration below)
I mean geometrically, changing depth will only move point along the ray (from a to b), but there is no way for it to "jump" to a different ray defined by another intrinsic matrix.
This is also a reason why I never felt Metric3D's scaling of focal lengths is sufficient
This is also a reason why I never felt Metric3D's scaling of focal lengths is sufficient
December 4, 2024 at 8:37 PM
I mean geometrically, changing depth will only move point along the ray (from a to b), but there is no way for it to "jump" to a different ray defined by another intrinsic matrix.
This is also a reason why I never felt Metric3D's scaling of focal lengths is sufficient
This is also a reason why I never felt Metric3D's scaling of focal lengths is sufficient
At the risk of appearing a fool - How is this true? It would mean that given two intrinsics K_0 and K_1 I could write:
dK_0^{-1}[u,v,1]^T = (ad + b)K_1^{-1}[u,v,1]^T
This is untrue as K_0 and K_1 will give us different rays. No change to the scalar will make the equation hold in general case?
dK_0^{-1}[u,v,1]^T = (ad + b)K_1^{-1}[u,v,1]^T
This is untrue as K_0 and K_1 will give us different rays. No change to the scalar will make the equation hold in general case?
December 4, 2024 at 6:28 PM
At the risk of appearing a fool - How is this true? It would mean that given two intrinsics K_0 and K_1 I could write:
dK_0^{-1}[u,v,1]^T = (ad + b)K_1^{-1}[u,v,1]^T
This is untrue as K_0 and K_1 will give us different rays. No change to the scalar will make the equation hold in general case?
dK_0^{-1}[u,v,1]^T = (ad + b)K_1^{-1}[u,v,1]^T
This is untrue as K_0 and K_1 will give us different rays. No change to the scalar will make the equation hold in general case?