Upload report/report.tex with huggingface_hub

report/report.tex

@@ -40,10 +40,9 @@ intermediate vertex detection stage.
 We introduce four technical contributions: (1)~\textbf{contrastive denoising
 training} that stabilises noisy Hungarian matching in early epochs; (2)~a
 \textbf{multi-scale encoder} that aggregates multiple encoder layer outputs
-via learned scalar weights; (3)~\textbf{Gestalt-guided point sampling} that
-upweights structurally important points at sample time without encoding them
-as model features; and (4)~\textbf{test-time augmentation} via four Y-axis
-rotations. Our model achieves a public test HSS of \textbf{0.575} (F1~=~0.664, IoU~=~0.516)
+via learned scalar weights; and (3)~\textbf{Gestalt-guided point
+sampling}~\cite{jastermark} that upweights structurally important points at
+sample time without encoding them as model features. Our model achieves a public test HSS of \textbf{0.575} (F1~=~0.664, IoU~=~0.516)
 and a best validation HSS of \textbf{0.534} on the cleaned val split.
 \end{abstract}
 
@@ -493,7 +492,7 @@ We presented WireframeDETR, a DETR-based end-to-end model for 3D building
 wireframe prediction that directly regresses edge coordinate pairs from a
 sparse RGB point cloud. Our four contributions — contrastive denoising
 training, multi-scale encoder memory, Gestalt-guided point sampling, and
-test-time augmentation — combine to achieve a public test HSS of \textbf{0.575}
+and Gestalt-guided point sampling — combine to achieve a public test HSS of \textbf{0.575}
 and a best validation HSS of 0.534
 on the S23DR 2026 cleaned val split. Post-processing routines are
 adopted from~\cite{jastermark}. Future work includes adding vertex feature