2/2
One of the generated videos. The point cloud is only used for building placement and for the topography, which was drastically altered in 1944. The CAD reconstruction is based on archival sources or on precise measurements (CGIS / photogrammetry) from the Plan et Reliefs.
#Brest1811
1/2
Still working on point-cloud densification to reconstruct #Brest1811. The method works: 7,500 images generated from 18 HD photos. Here’s an explanatory diagram of the workflow. In the next post, I’ll share an example video on the next post.
#Photogrammetry #RealityScan
#Brest1811 Modèle imprimé en 3D du Magasin Général, le guide de peinture que j’ai préparé, et le résultat simulé.
#Brest1811 Photo du Magasin général et reconstruction du fronton. Reconstruction également du fronton de la Grande Corderie. Un long travail à partir de photos du XIXᵉ siècle.
#RealityScan 2.1 finally fixes the bugs in v2.0.1 that was crashing five times a day and sometimes corrupting files. Updated yesterday and I can finally move forward! I’ve been struggling for six months with my hybrid reconstruction of #Brest1811, trying to work around those bugs.
#Brest1811
Reconstruction 3d de la batterie Royale et du quai aux victuailles toujours en cours en utilisant une technique mixte. Rendu pour l'étude.
#Brest1811 #RealityScan
Using AI video to densify point clouds? You'll get a nightmare of 70+ components that won't merge. We solved this by building an AI "Go-Player" (MCTS) that finds the perfect sequence. See our pipeline and GNN oracle path in the attached paper
drive.google.com/file/d/1UWAM...
Densifying RealityScan Point Clouds from AI-Generated Videos: Automatic, Optimized Multi-Component Alignment with a Go-Player Mindset We’ve been working on a pipeline to push RealityScan further than its usual “one scene = one component” workflow. The idea is to start from real photos, generate additional AI-based video frames to increase coverage, turn those into multiple RealityScan/RealityCapture components, and then automatically align and merge those components into a single, denser point cloud. The challenge is that RealityScan doesn’t always merge components in a predictable way: some pairs merge easily, others only merge if you do them earlier, and some will block you later if you play them too soon. That makes “just sort by score and merge” a poor strategy when you have a lot of small components coming from videos. “In practice, merging many RealityScan components is oddly similar to playing Go: every move changes the board, and a greedy move now can block a much better sequence later. That’s why we borrowed Monte Carlo Tree Search (originally popularized in computer Go) to explore better merge orders.” In the upcoming technical paper we’ll describe: • how we index frame folders and generate local RS batches (phase 1), • how we measure true pairwise mergeability inside RS and store it (phase 2), • and how we use an MCTS-guided orchestrator to pick the merge order that maximizes the final component size (phase 3). More soon.
#Brest1811
Densifying RealityScan Point Clouds from AI-Generated Videos: Automatic, Optimized Multi-Component Alignment with a Go-Player Mindset.
#RealityScan Point-cloud densification by adding AI-generated videos. Originally just 2–3 original photos. Visuals include: 1 “before” screen shot, then the result. (the integration requires a fairly complex workflow).
Here: one of the Recouvrance bastions.
#Brest #brest1811
#Brest1811 : reconstruction de la « Batterie Royale » et de « La Ninon » à partir de 4 photos et de 50 vidéos IA pour densifier le nuage de points. Le nuage de points est suffisant pour une modélisation CAO. « La Ninon » deviendra la première base de sous-marins à Brest en 1929.
WIP. After five years of research on #Brest1811, I can finally reconstruct areas once documented only by a few HD photos of the plan-relief. Using AI, I can generate a point cloud suitable for a precise manual reconstruction. Some of the areas reconstructed here were derived from just one photo.
Same CAD scene, two finishes:
- CAD-style, realistic graphic-novel render (Franco-Belgian BD)
- 18th-century French marine painting look (Vernet/Gudin)
Image order: Original 3D render → BD result → Painting result. #Brest1811 #3D #BD
Convict from the Brest penal colony, based on period engravings.
#Brest1811
I needed a break from #Brest1811.
I reworked @lasha3d.bsky.social artwork to make it 3D printable.
It was a big job on the files: adjusting thicknesses, adding supports, making the mesh watertight.
First results. need some rework.
I love his magical universe.
The vegetation will be added later.
#Brest #Brest1811
Ready for 3d printing with 3d textures.
Couvent des Carmes - ancienne église St Yves (rue St Yves maintenant rue Emile Zola).
#Brest #Brest1811. L’église Saint-Yves, ancienne église du couvent des Carmes, rue Saint-Yves (à peu près l’actuelle rue Émile-Zola, en contrebas de la place Wilson, ancien champ de bataille). Un véritable enfer à reconstruire à partir des archives et du plan-relief.
Pour mon projet #Brest1811, j’ai acheté les photos HD des plans-reliefs avec des droits spécifiques (sans droit de publication). Et je peux vous dire que ce n’est pas donné.
