{"paper":{"title":"Rethinking Graph Convolution for 2D-to-3D Hand Pose Lifting","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Adaptive attention outperforms fixed graph convolution for lifting 2D hand poses to 3D.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Chanyoung Kim, Donghyun Kim, Dong-Hyun Sim, Seong Jae Hwang, Youngjoong Kwon","submitted_at":"2026-05-13T14:39:24Z","abstract_excerpt":"Graph convolutional networks (GCNs) are widely used for 3D hand pose estimation, where the hand skeleton is encoded as a fixed adjacency graph. We revisit whether this is the most effective way to incorporate hand topology in 2D-to-3D lifting. In this paper, we perform controlled, parameter-matched ablations on the FPHA benchmark and show that standard multi-head self-attention consistently outperforms GCN baselines. Even when the GCN is strengthened with multi-hop adjacency and matched parameter count, self-attention reduces MPJPE from 12.36 mm to 10.09 mm. A skeleton-constrained graph attent"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"These results suggest that, for hand pose lifting, adaptive spatial attention is a more effective inductive bias than fixed graph convolution.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The assumption that parameter-matched ablations on the FPHA benchmark alone establish the general superiority of attention over GCNs across hand pose lifting tasks and datasets.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Self-attention with input-dependent aggregation and soft graph-distance priors outperforms fixed graph convolutions for 2D-to-3D hand pose estimation on FPHA.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Adaptive attention outperforms fixed graph convolution for lifting 2D hand poses to 3D.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"ecc16362c7c194e2361129087e9ad0e56b753880d1b6ad97573bcf9d35c7776e"},"source":{"id":"2605.13604","kind":"arxiv","version":1},"verdict":{"id":"e6ccb514-8b9b-4bf7-8902-ccad8de72b85","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:18:04.208070Z","strongest_claim":"These results suggest that, for hand pose lifting, adaptive spatial attention is a more effective inductive bias than fixed graph convolution.","one_line_summary":"Self-attention with input-dependent aggregation and soft graph-distance priors outperforms fixed graph convolutions for 2D-to-3D hand pose estimation on FPHA.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The assumption that parameter-matched ablations on the FPHA benchmark alone establish the general superiority of attention over GCNs across hand pose lifting tasks and datasets.","pith_extraction_headline":"Adaptive attention outperforms fixed graph convolution for lifting 2D hand poses to 3D."},"references":{"count":14,"sample":[{"doi":"","year":2018,"title":"First-person hand action bench- mark with rgb-d videos and 3d hand pose annotations","work_id":"4fcf3365-5d83-47df-8634-ef8ae945f996","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2017,"title":"Kipf and Max Welling","work_id":"5ad16f04-d1f4-40f0-a90d-8f54486aa7a3","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Gtignet: Global topology interaction graphormer network for 3d hand pose estimation.Neural Networks, 185:107221, 2025","work_id":"24ac1926-d62f-4442-a761-10d46a656967","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2019,"title":"Decoupled weight de- cay regularization","work_id":"f3b247f3-b210-4494-b445-0338427b4292","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Assemblyhands: Towards egocen- tric activity understanding via 3d hand pose estimation","work_id":"19eec250-2859-4986-a38c-c11fb2b1db68","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":14,"snapshot_sha256":"8303fbb739e907182f01eefd98113ac9462e0709e5bb6f519f237853f30f5d19","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"3832dece7c7ef158b4995356e1a24dd2d26d1bbde4710d1090e8a51717416f87"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}