Bruteforce matching, if you have plenty of time!

WebODM
#1

43 images of my bike shoes with my phone, thought I’d try Bruteforce matching, but wasn’t expecting it to take nearly a whole day vs 24 minutes!
It does extract a lot more features, but the time taken doesn’t justify it, as the resulting 3D models are similar.

43 images 21:51:29 Completed

|Options:|auto-boundary: true, dsm: true, feature-quality: ultra, matcher-neighbors: 10, matcher-type: bruteforce, mesh-octree-depth: 12, mesh-size: 300000, min-num-features: 12000, orthophoto-resolution: 0.05, pc-classify: true, pc-filter: 0, pc-geometric: true, pc-quality: ultra, resize-to: -1, use-3dmesh: true|
|Area:|133.76 m²|
|Reconstructed Points:|15,057,444|

43 images 00:24:20 Completed

|Options:|auto-boundary: true, dsm: true, feature-quality: ultra, matcher-neighbors: 10, mesh-octree-depth: 12, mesh-size: 300000, min-num-features: 12000, orthophoto-resolution: 0.05, pc-classify: true, pc-filter: 0, pc-geometric: true, pc-quality: ultra, resize-to: -1, use-3dmesh: true|
|Area:|149.8 m²|
|Reconstructed Points:|15,082,824|

Area is nonsense due to no GPS, so GSD set to default 5cm, when it was really more like <0.1mm

Matched 903 pairs (brown-brown: 903) in 77334.59451540001 seconds (85.64185439180511 seconds/pair).
2022-04-09 16:28:34,576 DEBUG: Good tracks: 169824

vs

Matched 903 pairs (brown-brown: 903) in 182.51813579999998 seconds (0.20212418172757474 seconds/pair).
2022-04-09 17:17:07,829 DEBUG: Good tracks: 61095

Bruteforce

Bike-shoes3D-Bruteforce
Bike-shoes3D-Bruteforce1631×912 185 KB

Default

Bike-shoes3D
Bike-shoes3D1614×892 196 KB

4 Likes
3D Model using Cell Phone
#2

Nice models!

2 Likes
#3

I am wondering how you are getting these images. I tried with the following (image) and it seems to build an entire world out there. While ignoring the object of interest.

IMG_20220507_113053
IMG_20220507_1130531920×2560 524 KB

1 Like
#4

Shiny surfaces are very difficult to reconstruct, especially with saturated areas of the image, plus the reflections, which change depending on where you are photographing from.

3 Likes
#5

Is this also relevant to statues? I was planning to take some images in the neightbourhood and reconstruct them using OpenODM.

#6

Shiny surfaces are a problem no matter where they are. Statues of non-reflective materials should reproduce fairly well with a sufficient number of images, so long as you can exclude most of any noisy backgrounds and sky, or else mask them out.
Using a boundary to restrict the area of reconstruction will help a lot too.

2 Likes
#7

Ok. I’ll try with another subject.
A quick question - while taking the initial photos, should the entire subject be part of the photo or it’s okay if I take only parts of it in the image.

2 Likes
#8

You don’t have to have any photos that show the whole subject, but you generally do need enough overlap and photos from multiple angles covering each point to get a good result. For non-uniform surfaces like statues that may include occluded areas, it can be a bit tricky, requiring more photos than you might expect.

You will probably have to adjust your parameters by trial and error a bit to get a good result, such as I did here -

1 Like