<html><body><div style="font-family: arial, helvetica, sans-serif; font-size: 12pt; color: #000000"><div><div>Dear colleagues,</div><div><br data-mce-bogus="1"></div><div>The IGN/LASTIG lab and SIRADEL company propose a PhD position on <strong>Large-scale 3D modeling of LOD2 buildings using Deep Learning</strong></div><div><a data-mce-href="https://www.umr-lastig.fr/lastig_data/pdf/2024_sujet_th%C3%A8se_Cifre_LOD2_Siradel_IGN_light_en.pdf" href="https://www.umr-lastig.fr/lastig_data/pdf/2024_sujet_th%C3%A8se_Cifre_LOD2_Siradel_IGN_light_en.pdf">Full description</a></div><div><br data-mce-bogus="1"></div><div><strong>Keywords</strong> : Digital Twin, 3D Reconstruction, LOD2 Modeling, Deep Learning, Segmentation</div><div><strong><br data-mce-bogus="1"></strong></div><div><strong>Supervision</strong> :</div><div>•Director: Bruno Vallet (LASTIG, ENSG, IGN)</div><div>•Supervisers :</div><div>oJérôme Dolbecq (Département Territoires Numériques, SIRADEL - ENGIE)</div><div>oGuillaume Despine (Département Territoires Numériques, SIRADEL - ENGIE)</div><div>oWu Teng (LASTIG, ENSG, IGN)</div><div><strong><br data-mce-bogus="1"></strong></div><div><strong>Localisation</strong> :</div><div>•Year 1 : ACTE Team - LASTIG Lab – Champs sur Marne (77) FRANCE</div><div>•Years 2 and 3 : RD DT Team - Siradel – ENGIE – Rennes (35) FRANCE</div><div>•Occasional travel between the two sites will be expected.</div><div><strong><br data-mce-bogus="1"></strong></div><div><strong>Financing</strong> : CIFRE (industrial) PhD</div><div><br data-mce-bogus="1"></div><div><strong>Goals:</strong></div><div><div style="">The aim of this PhD is to automate LOD2 modeling from airborne data covering large geographical areas. The source data can be LiDAR point clouds (10-15 points/m²) or Digital Surface Models (Stereo correlation - 20-50cm resolution) which will have been automatically classified.</div><div style="">The algorithm developed will be based on Deep Learning. End-to-End methods seem promising because they avoid the propagation of errors between the different stages. However, the chosen solution must be pragmatic and take into account the reality of industrial production (noise, incompleteness, classification error). The topology of the 3D models will also be at the heart of concerns (watertightness, manifoldness, flatness). The neural network must therefore be able to qualify the input data, to assess to what extent it can trust it, and also evaluate the 3D model produced by indicating a reliability score that can be interpreted by a non-expert.</div><div style="">The developed algorithm will first have to confront the Building3D challenge, then Siradel production data which presents a wide variety of urban and architectural configurations.</div><div style="">In urban areas, the density of buildings forms large blocks of connected buildings. The individualization of these buildings is generally little addressed in studies, which consider the 2D contours always available. From an industrial point of view, however, this is a significant stake. Automatic segmentation of building instances (region growth, super-points, etc.), integrated into the processing chain, would therefore be a major additional contribution.</div><div style=""><br data-mce-bogus="1"></div><div style=""><div><strong>Application:</strong></div><div>Send an email to bruno.vallet@ign.fr and sbenitez@siradel.com with:</div><div>•Your resume</div><div>•A cover letter</div><div>•A transcript of your master’s grades</div><div>•One or more letters of recommendation (if applicable)</div></div></div><div><strong><br data-mce-bogus="1"></strong></div></div></div></body></html>