Mar 30
Very cool work! We followed a quite similar pipeline a couple of year ago to generate scenes from a single-exemplar: arxiv.org/abs/2412.16253. In our case, latent features were DINO and Gaussians were decoded with a patch-based synthesis step (inspired by PatchMatch).
1
7
1,222
25 Nov 2025
MetroGS: Efficient and Stable Reconstruction of Geometrically Accurate High-Fidelity Large-Scale Scenes
Kehua Chen, et al.
tl;dr: 2DGS dense enhancement with π3 monocular&PatchMatch-based multi-view opt. depth-guided appearance modeling with Tri-MipRF
arxiv.org/abs/2511.19172
4
38
1,626
17 Nov 2025
昔はこんな感じでしたよね。
2009年にAdobeのPatchMatchの論文が発表されて、Photoshopに「コンテンツに応じた塗りつぶし」などの機能としてある程度取り入れられたときには、衝撃を受けましたが、
それでも今の生成AIでの編集に比べるとオモチャレベルでしたね。
今の生成AIと違い、周囲で似たピクセルを探すようなアルゴリズムで実現していました。
1
9
17,256
30 Aug 2025
ふとしたことから、PatchMatchアルゴリズムによるMulti-View Stereoをフィックスターズ社がCUDA高速化( docswell.com/s/fixstars/KV1V… )したCUMVS( github.com/fixstars/cuda-mul… )の存在を知りました。Apache License 2.0ですし「処理速度を飛躍的に向上」させたとのことで、興味深く。
4
431
22 Jul 2025
👉👉 MVP-Stereo: A #Parallel Multi-View Patchmatch #Stereo Method with #Dilation Matching for #Photogrammetric Application
✍️ Qingsong Yan et al.
🔗 brnw.ch/21wUkhi
2
11
499
3 Oct 2024
I think we needed mature non-learned systems to generate training data for Dust3r and others. A lot of the datasets used for training are generated with SLAM, SfM, PatchMatch MVS, etc.
1
9
1,626
27 Aug 2024
That is patchmatch based though? You dont get the precision of traditional methods with current learning based, but the fusion step actually can solve most issues.
My original statement still stands though ;)
1
3
302
14 Aug 2024
【セミナー資料】8/7開催のコンピュータビジョンセミナー「3次元復元アルゴリズム Multi View StereoのCUDA高速化」の当日資料を公開しました。
PatchMatch MVSのアルゴリズムと、GPUによる並列化の手法を解説しています。次回もぜひご期待ください！
news.fixstars.com/4841/
1
8
689
8 Aug 2024
While being great, this is a license centipede 😅. The repo is MIT, which uses agpl licensed code, which implements ebsynth, which uses an Adobe patented patchmatch Algo. 🥲
2
5
532
23 Jul 2024
【セミナー】コンピュータビジョンセミナー「3次元復元アルゴリズム Multi View StereoのCUDA高速化」を開催します→8/7(水)18:00～無料
PatchMatch MVSのアルゴリズムと、GPUによる並列化の手法を解説予定。コンピュータビジョン関連のエンジニア、学生の皆様におすすめです!
news.fixstars.com/4841/
9
22
2,703
23 May 2024
PatchMatch was pure magic when it was introduced 14 years ago. It allowed you to shrink images, delete and even move objects around. It eventually turned into Photoshop content-aware fill.
1
95
30 Apr 2024
Some optical flow works. Also patchmatch kind of works. The main cool thing is that there's so many depth borders. I really like it.
1
3
140
29 Apr 2024
I have very little intuition for what matters in patchmatch actually, but a very big distribution shift between proposals and actual seems like it should matter.
1
95
4 Apr 2024
Another thing that I found quite useful in improving patchmatch output is this refinement approach: github.com/rossimattia/depth…. You might give it a shot.
1
5
321
4 Apr 2024
Here's an example, I think you can see some artifacts of patchmatch. For example that you tend to get "constant depth" regions, that can bleed outside the object boundaries.
3
3
263
4 Apr 2024
To be fair with patchmatch, there have been significant improvement over colmap impl (e.g. ieeexplore.ieee.org/document…). In my own tests, it is consistently better. I think a good contribution to the community would be to port the code to colmap in terms of style and UX/UI.
1
2
123
4 Apr 2024
I should note that patchmatch generally works well, but I feel like there should have been some development the last 15 years.
4
182
4 Apr 2024
Can someone pls make better MVS than patchmatch, I'm begging you. Otherwise I'm making it this autumn.
6
18
4,040