Multimodal Representation Learning for 3D Model RetrievalČatar, Neli (Avtor) Skočaj, Danijel (Mentor) Miščič, Andrej (Komentor) Vranješ, Luka (Komentor) 3D model retrievalmultimodal learningcontrastive learningCLIPmulti-view renderingjoint embedding spaceattention pooling3D representationsRetrieving 3D models from large industrial repositories is a challenging problem, particularly when queries take the form of natural language descriptions or other 3D shapes. Existing approaches typically address text-to-3D and 3D-to-3D retrieval as separate tasks, limiting their practical utility in settings where both capabilities are required simultaneously. This thesis presents a unified multimodal retrieval framework that supports both query types within a single shared embedding space. 3D objects are represented as sets of greyscale rendered views, focusing on geometric shape rather than colour or texture. A joint training objective combines cross-modal contrastive learning for text-to-3D alignment with an auxiliary render-based objective for 3D-to-3D consistency. To support large-scale training, a dataset of 537,000 shape-annotated objects was constructed by automatically generating text descriptions from the Objaverse-XL dataset using the Qwen3 32B vision language model. The best-performing model utilised attention pooling, together with the debiased contrastive loss and a batch size of 80. The backbone was laion/CLIP-ViT-L-14-laion2B-s32B-b82K. On text-to-3D retrieval, the model achieved an R@20 of 0.7297, while on 3D-to-3D retrieval, it achieved an R@20 of 0.9968. The framework is further evaluated on ModelNet40 and Text2Shape, and an exploratory image-to-3D adapter is investigated as an extension toward image-query support.20262026-10-23 17:15:04Magistrsko delo190189VisID: 38522sl