22 
In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C.. Tournaire O. (Eds), 1APRS. Vol. XXXVIII. Part 3A - Saint-Mandé, France. September 1-3. 2010 
Figure 1: Geometry estimation steps: (a) day 1 sequence, (b) re 
move drift, (c) merge day 1-2 sequences, (d-f) remove drifts, (g) 
use all features. All results are registered in rectangle [0.1] x 
[0. 0.8] by enforcing constant coordinates on the two poses sur 
rounded by gray disks in (a). Gray disks in (b,d,e,f) show poses 
where drift is corrected. Day 1-2 sequences are merged on gray 
disk in (c). 
remove drifts using k — 1. Cases (b,d,e.f) of Fig. 1 are trajec 
tory loops with (424,451,1434,216) images and are obtained by 
(16.62,39,9) CBA iterations in (190,2400.1460,370) seconds, re 
spectively. We think that a large part of the drift in case (d) is 
due to the single view point approximation, which is inaccurate 
in the outdoor corridor (top right corner of Fig. 4) with small 
scene depth. A last BA is applied to refine the geometry (3D 
and intrinsic parameters) and to increase the list of reconstructed 
points. The final geometry (Fig. l.g) has 699410 points recon- 
staicted from 3.9M Harris features; the means of track lengths 
and 3D points visible in one view are 5.5 and 1721. respectively. 
Then, 2256 view-centered models are reconstructed thanks to the 
methods in Section 2.4 and 2.5 using k — 1. This is the most time 
consuming part of the method since one view-centered model is 
computed in about 3 min 30s. The first step of view-centered 
model computation is the over-segmentation mesh in the refer 
ence image. It samples the view field such that the super-pixels at 
the neighborhood of horizontal plane projection are initialized by 
squares of size 8x8 pixels in the images. The mean of number 
of 3D triangles is 17547. Fig. 3 shows super-pixels of a reference 
image and the resulting view-centered model. 
Figure 2: From top to bottom: 722 and 535 matches of L t ,j used 
to remove drift in cases (d) and (e) of Fig. 1. Images of days 1 
and 2 are on the left and right, respectively. 
Last, the methods in Section 2.6 are applied to filter the 39.6M 
triangles stored in hard disk. A first filtering is done using relia 
bility («o = 5 degrees), prior knowledge and uncertainty filters 
(wo = 1-1): we obtain 6.5M triangles in 40 min and store them in 
RAM. Redundancy removal is the last filtering and selects 4.5M 
triangles in 44 min. Texture packing and VRML file saving take 
9 min. Fig. 4 shows views of the final model. We note that the 
scene is curved as if it lie on a sphere surface whose diameter has 
several kilometers: a vertical component of drift is left. 
An other experiment is the quantitative evaluation of scene ac 
curacy (discrepancy between scene reconstruction and ground 
truth) for a view-centered model using k = 1. A represen 
tative range of baselines is obtained with the following ground 
truth: the [0, 5] 3 cube and camera locations defined by c, = 
(l 1 + ?'/5 l) ,i € {0,1, 2} (numbers in meters). First, 
synthetic images are generated using ray-tracing and the knowl 
edge of mirror/perspective camera/textured cube. Second, meth 
ods in Sections 2.1, 2.2. 2.4 and 2.5 are applied. Third, a camera- 
based registration is applied to put the scene estimation in the 
coordinate frame of ground tnith. Last, the scene accuracy ao.9 
is estimated using the distance e between vertex v of the model 
and the ground truth surface: inequality |e(v)| < ao.g||v — ci|| 
is true for 90% of vertices. We obtain 00.9 = 0.015. 
4 CONCLUSION 
We present an environment reconstruction system from images 
acquired by a $1000 camera. Several items are described: camera 
model, structure-from-motion. drift removal, view field sampling 
by super-pixels, view-centered model and triangle filtering. Un 
like previous work, image meshes define both super-pixels (con 
vex polygons) and triangles of 3D models. The current system 
is fully automatic up to the loop detection step (that previous 
methods could solve). Last it is experimented on a challenging 
sequence. 
Future work includes loop detection integration, better use of 
visibility and prior knowledge for scene reconstruction, joining