BEGIN:VCALENDAR VERSION:2.0 PRODID:-//wp-events-plugin.com//7.4.0.1//EN TZID:Asia/Kolkata X-WR-TIMEZONE:Asia/Kolkata BEGIN:VEVENT UID:225@cds.iisc.ac.in DTSTART;TZID=Asia/Kolkata:20260925T110000 DTEND;TZID=Asia/Kolkata:20260925T120000 DTSTAMP:20260903T072204Z URL:https://cds.iisc.ac.in/events/ph-d-thesis-colloquium-102-cds-25-septem ber-2026-from-reconstructing-scenes-to-geometry-guided-synthesis/ SUMMARY:Ph.D: Thesis Colloquium: 102: CDS: 25\, September 2026 "From Recons tructing Scenes to Geometry-Guided Synthesis" DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nPh.D. Thesis Col loquium\n\n\n\nSpeaker: Mr. Ankit Dhiman\nS.R. Number: 06-18-00-11-12-21-1 -20342\nTitle: "From Reconstructing Scenes to Geometry-Guided Synthesis"\n Research Supervisor: Prof. Venkatesh Babu\nDate &\; Time : September 25 \, 2026 (Friday)\, 11:00 AM\nVenue : #102\, CDS Seminar Hall\n\n\n\nABSTRA CT\nRecent 3D representations such as Neural Radiance Fields and 3D Gaussi an Splatting have advanced the field of photorealistic novel-view synthesi s from posed images. Beyond the task of novel-view synthesis\, these 3D re presentations aggregate 2D foundation model predictions across views and p rovide geometric conditional signals to generative models for complex task s. This thesis explores this broader role\, progressing from (1) improving the underlying 3D representations\, to (2) resolving inconsistencies in c ross-view predictions\, and finally (3) using recovered geometry to guide generative models. Specifically\, the thesis is organized into three parts :\n\n Neural Scene Representations: The first part focuses on three limit ations of current neural scene representations: (i) modeling stratified sc enes\, (ii) the high computational cost for reconstruction form high-resol ution multi-view images\, and (iii) aliasing in dynamic scenes. Strata-NeR F models stratified scenes using a vector-quantized latent-conditioned rad iance field that learns transitions between different levels without super vision. Turbo-GS accelerates the fitting of 3D Gaussian Splatting for high -resolution multi-view inputs. For dynamic scenes\, we show that motion ch anges the effective sampling rate and derive a motion-aware filter to redu ce aliasing across scale and time. Together\, these works improve the effi ciency and robustness of neural scene representations.\n Lifting 2D Cues into 3D: The second part uses a 3D representation to aggregate 2D predicti ons that are consistent for a single view but inconsistent across views. C hromaDistill colorizes scenes captured without color (legacy grayscale and Infrared image sequences) by distilling a pretrained colorization model i nto the 3D representation during training. This makes the predicted color consistent across views. UniC-Lift lifts per-image instance masks into a 3 D Gaussian representation and directly predicts consistent instance labels \, avoiding the post-processing clustering stage used by earlier methods.\ n Geometry-Conditioned Generation: The third part explores whether explic it scene geometry can constrain generative models. Mirror reflection synth esis is a challenging test for generative models as the reflected content on the mirror should be determined by the scene geometry and an incorrect reflection is immediately apparent. Reflecting Reality introduces a depth- conditioned approach to the aforementioned task together with the large-sc ale synthetic SynMirror dataset. MirrorVerse improves its generalization b y introducing greater diversity in synthetic scenes through SynMirrorV2. G eoMirror takes a further step by computing the reflected scene geometry ex plicitly and using it to guide generation. Finally\, GeoNoise extends this idea beyond reflections\, using scene geometry to guide the generative pr ocess itself for training-free\, geometry-consistent novel-view synthesis. \n\nThis thesis advances neural scene representations by making them more efficient and robust to aliasing artifacts\, and shows the importance of m aking geometry explicit for reliable visual prediction and generation. Thi s is important when correctness is defined in physical terms\, and plausib ility is insufficient. As geometry-consistent novel-view synthesis and hig h-fidelity dynamic reconstruction continue to advance\, these ideas open n ew avenues toward monocular dynamic novel-view synthesis and geometrically consistent video and world models.\n\n\n\nALL ARE WELCOME CATEGORIES:Events,Ph.D. Thesis Colloquium END:VEVENT BEGIN:VTIMEZONE TZID:Asia/Kolkata X-LIC-LOCATION:Asia/Kolkata BEGIN:STANDARD DTSTART:20250925T110000 TZOFFSETFROM:+0530 TZOFFSETTO:+0530 TZNAME:IST END:STANDARD END:VTIMEZONE END:VCALENDAR