Jordi Gonzalez

M 48 ✗ Ours ✓ ✓ ✓ 80 3,273 474 72., D sewing pattern ([9]), which provides clothing’s direct measurements. This is not limited to garment images; clothed humans [46, 44] and even text guidance [5] can also be used. Taking all that into account, we leverage our data to predict garment size measurements directly from images of clothed individuals. Our approach adapts SPnet [45], a recent model designed to predict sewing parameters from posed people, in which an encoder-decoder network Ψ​(Gs,Ps,Pt)\Psi(G^{s},P^{s},P^{t}) is used to predict the target garment normal image GtG^{t} in a canonical pose, where GsG^{s}, PsP^{s} and PtP^{t} are image representations for the input garment normal, input body pose, and output canonical pose, respectively. Then, a separate network Φ​(Gt)\Phi(G^{t}) is trained to estimate the sewing parameters. We refer the reader to [45] for the original architecture details. Given motion sequences, we obtain GsG^{s} with our segmentation and Sapiens [37], and PsP^{s} from our SMPL-X estimation. Likewise, GtG^{t} and PtP^{t} are computed from the template recordings. This allows us to train on our real data rather than the synthetic one used in SPnet. We follow the same architecture and training protocol as SPnet for the network Ψ\Psi. However, we update Φ\Phi to regress normalized garment sizes, scaled to [0,1][0,1] via dividing by the maximum size in the dataset. We evaluate three variants of Φ\Phi: (1) individual models trained per garment group (Sec. 3.3), as in SPnet; (2) a single multi-task model trained jointly on all groups; and (3) a multi-task model where the SegNet [3] encoder in (2) is replaced with SwinV2 [49] and additionally conditioned on PtP^{t}, to better align garment regions with the corresponding sizing chart entries. 4.3 Novel View Synthesis (NVS) Novel View Synthesis is a critical component of modern virtual try-on and digital fashion applications, enabling the creation of realistic, 360-degree views of clothed individuals. Recent state-of-the-art methods for clothed human reconstruction [42], garment generation [67], and