S2HAND: Accuracy-Preservation Trade-Offs for Hand Specialization in Whole-Body Pose Models
Organizations: Independent Researcher · Tsinghua University · Japan Advanced Institute of Science and Technology
Abstract
Unified whole-body pose estimators jointly predict body, face, and hand landmarks, but often degrade on small, blurred, or occluded hands. Attaching a standalone hand specialist introduces extra detection overhead, coordinate alignment, and wrist-stitching artifacts. We introduce S2HAND, a hand-specialized Sapiens2-5B framework that preserves the native 308-keypoint output interface. Specifically, S2HAND-MT adapts a lightweight 0.87M-parameter pose branch to route 40 finger heatmaps while strictly preserving the shared Transformer, auxiliary task heads, and the other 268 body, foot, and face heatmaps. S2HAND-Pose further fine-tunes three Fisher-selected Transformer blocks (267.13M parameters), trading exact preservation for stronger hand recovery: it lifts macro AUC from 57.83% to 65.73% across HInt TEST, COCO WholeBody VAL, and FreiHAND EVAL, and surpasses HInt-adapted Hand5/HRNet and stitched Hand5 systems on all five Native Pipeline benchmarks. On COCO, S2HAND-Pose improves hand AP by 12.89 points and body-plus-foot AP by 0.19 points. Across a 512-image multi-task audit, it maintains 99.60% foreground segmentation agreement with the public checkpoint, with quantified normal and pointmap drift. Its 2D coordinates also lower 3D mesh error in frozen Pose2Mesh transfer, while our optional TailGuard re-cropping module cuts S2HAND-MT's severe failure rate on NewDays from 13.45% to 10.43%. S2HAND thus resolves the whole-body hand bottleneck with either exact non-hand preservation (MT) or empirically audited trade-offs (Pose).
Figures & tables
| System | NewDays (TEST) | VISOR (TEST) | H2O (TEST) | EPICK (VAL) | H2SMV-FPV (synthetic TEST) |
| Public Sapiens2-5B [ 20 ] | 61.48 / 14.14 | 56.12 / 17.68 | 69.02 / 4.38 | 55.16 / 18.56 | 47.93 / 29.86 |
| RTMPose-Hand5 (public) [ 15 ] | 59.72 / 16.59 | 63.51 / 10.86 | 67.43 / 0.43 | 64.10 / 8.48 | 40.57 / 39.00 |
| RTMPose-Hand5 + HInt FT [ 15 ] | 61.57 / 17.33 | 69.85 / 10.02 | 72.57 / 0.30 | 71.01 / 6.40 | 39.78 / 39.64 |
| FreiHAND-HRNet reproduction [ 32 ] | 17.62 / 79.19 | 28.32 / 54.09 | 60.19 / 7.29 | 27.84 / 55.20 | 26.66 / 61.59 |
| FreiHAND-HRNet + HInt FT [ 32 ] | 39.42 / 40.31 | 56.57 / 18.57 | 68.74 / 1.80 | 57.14 / 15.20 | 35.27 / 47.49 |
| EffHandEgoNet-H2O [ 28 ] | 5.20 / 92.30 | 32.25 / 42.55 | 75.25 / 0.06 | 29.11 / 45.60 | 8.37 / 84.97 |
| System | NewDays (TEST) | VISOR (TEST) |
|---|---|---|
| RTMPose-Hand5 (public) [ 15 ] | 57.30 | 59.19 |
| RTMPose-Hand5 + HInt FT [ 15 ] | 60.21 | 65.86 |
| FreiHAND-HRNet + HInt FT [ 32 ] | 40.18 | 49.43 |
| S2HAND-Pose (Ours) | 74.27 | 78.34 |
| External Hand/Wrist Policy | NewDays | VISOR | H2O | EPICK | H2SMV-FPV |
|---|---|---|---|---|---|
| RTMPose-Hand5 (public) / replace all | 63.11 | 65.88 | 67.45 | 66.23 | 47.85 |
| RTMPose-Hand5 (public) / keep wrist | 62.82 | 65.73 | 67.04 | 66.10 | 47.86 |
| RTMPose-Hand5 (public) / translate fingers | 36.72 | 44.50 | 56.00 | 44.08 | 35.49 |
| RTMPose-Hand5 + HInt FT / replace all | 64.96 | 73.45 | 72.59 | 73.14 | 47.06 |
| RTMPose-Hand5 + HInt FT / keep wrist | 64.41 | 73.14 | 71.05 | 72.00 | 47.22 |
| RTMPose-Hand5 + HInt FT / translate fingers | 34.97 | 56.18 | 40.39 | 40.38 | 29.52 |
| Adaptation Scope | Trainable Params (M) | Model Updated (%) | HInt VAL [email protected] | Mean (3 sets) [email protected] | Face mean NME | Face [email protected] (%) | Preservation |
|---|---|---|---|---|---|---|---|
| Public Sapiens2-5B | 0 | 0 | – | 57.83 | 0.1428 | 49.25 | public reference |
| Rank-64 adapter | 0.033 | 0.0006 | 67.49 | 58.07 | – | – | empirical |
| Finger-output rows | 0.079 | 0.0015 | 67.49 | 58.08 | – | – | channel-local |
| Hand-output rows | 0.079 | 0.0015 | 67.49 | 58.08 | – | – | channel-local |
| Decoder tail | 0.210 | 0.0041 | 67.64 | 58.16 | – | – | empirical |
| Ordinary pose head | 0.867 | 0.0169 | 68.47 | 58.92 | – | – | empirical |
| Endpoint | Public Sapiens2-5B | S2HAND-MT (Ours) | S2HAND-Pose (Ours) |
|---|---|---|---|
| Protected pose heatmaps | reference | exact 268-channel route | no exact guarantee |
| COCO body+foot AP / AR | 79.09 / 82.83 | public route | 79.28 / 83.07 |
| COCO hand AP / AR | 51.92 / 57.18 | 54.75 / 60.41 | 64.81 / 70.38 |
| Face 16-anchor mean NME | 0.1428 | public route | 0.1510 |
| Face 16-anchor [email protected] | 49.25% | public route | 53.70% |
| Segmentation foreground agreement | reference | unchanged path | 99.5956% |
| 2D Driver | FreiHAND | H2O | DexYCB | AssemblyHands | H2SMV-FPV |
|---|---|---|---|---|---|
| Public Sapiens2-5B [ 20 ] | 11.746 | 9.923 | 13.400 | 18.470 | 14.597 |
| RTMPose-Hand5 (public) [ 15 ] | 8.768 | 9.705 | 9.931 | 18.043 | 16.360 |
| S2HAND-MT (Ours) | 11.306 | 9.917 | 13.151 | 18.104 | 14.388 |
| S2HAND-Pose (Ours) | 8.424 | 8.598 | 10.783 | 16.025 | 12.521 |
| GT 2D (reference) | 4.772 | 5.901 | 5.522 | 17.080 | 7.888 |
| Public Pose [95% CI] |
| Hand XY Source | Top-1 | Macro Top-1 |
| EffHandEgoNet-H2O | 95.90 | 95.60 |
| Public Sapiens2-5B | 94.26 | 93.75 |
| S2HAND-MT (Ours) | 92.62 | 91.90 |
| S2HAND-Pose (Ours) | 95.08 | 94.68 |
| GT 2D (reference) | 96.72 | 96.30 |
| Zero hands (control) | 72.95 | 70.37 |
| System | Total Params | Adapted Params | Latency (ms) | Boundary |
|---|---|---|---|---|
| Public Sapiens2-5B [ 20 ] | 5.121B | 0 | 320.50 | forward |
| RTMPose-Hand5 (public) [ 15 ] | 13.775M | 0 | 4.61 | hand forward |
| EffHandEgoNet-H2O [ 28 ] | 25.242M | 0 | 10.10 | native forward |
| Public + RTMPose-Hand5 (public) | 5.161B | 0 | component estimate | |
| Public + RTMPose-Hand5 + HInt FT | 5.161B | 13.775M | component estimate | |
| Public + WiLoR [ 31 ] | 5.787B | 0 | 356.07 | component estimate |
Appendix figures & tables46 assets
Supplementary material from the paper’s appendix.
Appendix
| Dataset | Domain/Annotation | Role | Resampling Unit |
|---|---|---|---|
| HInt-NewDays TRAIN [ 30 ] | difficult visibility-aware real hands | training; balanced 1:1 with EPICK TRAIN | video/sequence |
| HInt-EPICK TRAIN [ 30 ] | difficult visibility-aware real hands | training; balanced 1:1 with NewDays TRAIN | video/sequence |
| HInt-NewDays VAL [ 30 ] | difficult visibility-aware real hands | checkpoint selection; macro-averaged with EPICK VAL | video/sequence |
| HInt-EPICK VAL [ 30 ] | difficult visibility-aware real hands | checkpoint selection and TailGuard selection | video/sequence |
| HInt-NewDays TEST [ 30 ] | difficult visibility-aware real hands | final HInt test endpoint | video/sequence |
| HInt-VISOR TEST [ 30 ] | difficult visibility-aware real hands | final HInt test endpoint | video/sequence |
| Dataset | System | Hands | [email protected] | [email protected] | [email protected] | [email protected] | Tail .20 |
|---|---|---|---|---|---|---|---|
| NewDays TEST [ 30 ] | Public Sapiens2-5B [ 20 ] | 1754 | 61.48 | 47.72 | 70.69 | 81.40 | 14.14 |
| RTMPose-Hand5 (public) [ 15 ] | 1754 | 59.72 | 48.63 | 69.04 | 77.46 | 16.59 | |
| RTMPose-Hand5 + HInt FT [ 15 ] | 1754 | 61.57 | 52.82 | 70.49 | 77.62 | 17.33 | |
| FreiHAND-HRNet reproduction [ 32 , 31 ] | 1754 | 17.62 | 8.66 | 17.55 | 26.68 | 79.19 | |
| FreiHAND-HRNet + HInt FT [ 32 ] | 1754 | 39.42 | 27.47 | 43.68 | 54.26 | 40.31 | |
| EffHandEgoNet-H2O [ 28 ] | 1754 | 5.20 | 1.57 | 4.94 | 8.57 | 92.30 |
| NewDays (TEST) | VISOR (TEST) | |||||
| System/Source | [email protected] | [email protected] | [email protected] | [email protected] | [email protected] | [email protected] |
| HaMeR + HInt (published) [ 30 , 5 ] | 51.6 | 81.9 | 91.9 | 56.5 | 88.1 | 95.6 |
| HandOS-2D (published) [ 5 ] | 55.8 | 75.8 | 84.5 | 66.2 | 85.3 | 91.8 |
| HandOS-proj (published) [ 5 ] | 53.7 | 75.9 | 85.1 | 64.8 | 85.4 | 92.0 |
| HandOS DINO-X, 2D (published) [ 5 ] | – | 76.3 | – | – | 84.8 | – |
| HandOS DINO-X, proj (published) [ 5 ] | – | 76.5 | – | – | 84.6 | – |
| Model | H2O (TEST) [ 22 ] | NewDays (TEST) [ 30 ] | EPICK (VAL) [ 30 ] | H2SMV-FPV (TEST) | FreiHAND (EVAL) [ 41 ] | Mean (4 sets) | Mean (5 sets) |
| Public Sapiens2-5B [ 20 ] | 70.53 | 65.69 | 65.73 | 51.96 | 60.62 | 63.48 | 62.91 |
| RTMPose-Hand5 (public) [ 15 ] | 66.64 | 57.30 | 61.98 | 42.97 | 65.59 | 57.22 | 58.90 |
| RTMPose-Hand5 + HInt FT [ 15 ] | – | 60.21 | – | – | – | – | – |
| FreiHAND-HRNet + HInt FT [ 32 ] | – | 40.18 | – | – | – | – | – |
| EffHandEgoNet-H2O [ 28 ] | 38.90 | 13.04 | 23.64 | 11.35 | 12.10 | 21.73 | 19.81 |
| COCO-Hand HRNet (official) [ 32 ] | 62.41 | 56.44 | 60.37 | 38.99 | 57.40 | 54.55 | 55.12 |
| Dataset | Model | GT Hands | Coverage (%) | [email protected] (%) | [email protected]/.10/.15 (%) | Side Swap (%) | Tail .20 (%) |
|---|---|---|---|---|---|---|---|
| H2O TEST subset [ 22 ] | Public Sapiens2-5B [ 20 ] | 4093 | 99.93 | 71.20 | 52.70/86.48/96.46 | 0.00 | 0.61 |
| RTMPose-Hand5 (public) [ 15 ] | 4093 | 99.93 | 52.49 | 33.35/63.62/74.14 | 17.81 | 26.46 | |
| FreiHAND-HRNet reproduction [ 32 ] | 4093 | 99.93 | 57.17 | 40.90/68.71/77.87 | 10.67 | 20.35 | |
| EffHandEgoNet-H2O [ 28 ] | 4093 | 99.93 | 73.94 | 61.58 /88.39/95.83 | 0.00 | 0.42 | |
| S2HAND-MT (Ours) | 4093 | 99.93 | 72.16 | 53.77/88.63/97.06 | 0.00 | 0.59 | |
| S2HAND-Pose (Ours) | 4093 | 99.93 | 74.80 | 60.11/ 91.70 / 98.09 | 0.00 | 0.10 |
| Setting | RTMPose-Hand5 + HInt FT | FreiHAND-HRNet + HInt FT |
|---|---|---|
| Initialization | public Hand5 | reproduced FreiHAND-HRNet-W18 |
| Optimizer | AdamW, decay 0.01 | Adam |
| Peak LR / final LR | / | / |
| Schedule | 800-sample warm-up, cosine | 800-sample warm-up, cosine |
| Input / bbox padding | / 1.25 | / 1.25 |
| BBox scale / rotate-factor | / | / |
| External Model | Dataset | Coverage (%) | Gap Median | Gap P95 | Gap P99 |
|---|---|---|---|---|---|
| RTMPose-Hand5 (public) | NewDays TEST | 91.16 | 0.102 | 1.169 | 2.991 |
| RTMPose-Hand5 (public) | VISOR TEST | 92.50 | 0.069 | 2.508 | 4.193 |
| RTMPose-Hand5 (public) | H2O TEST | 99.96 | 0.047 | 0.152 | 3.029 |
| RTMPose-Hand5 (public) | EPICK VAL | 95.04 | 0.073 | 2.912 | 3.862 |
| RTMPose-Hand5 (public) | H2SMV-FPV TEST | 71.16 | 0.072 | 1.439 | 2.701 |
| RTMPose-Hand5 + HInt FT | NewDays TEST | 91.16 | 0.114 | 1.135 | 2.995 |
| Stage | Parameters | Median (ms) | Mean (ms) |
|---|---|---|---|
| WiLoR-YOLO detector | 26.631M | 5.94 | 6.63 |
| RTMPose-Hand5 forward (one hand) | 13.775M | 4.61 | – |
| WiLoR detector + regressor | 665.912M | 35.57 | 50.70 |
| Public + Hand5 component sum | 5.161B | – | |
| Public + WiLoR component sum | 5.787B | 356.07 | – |
| Model | NewDays (TEST) | VISOR (TEST) | H2O (TEST) | EPICK (VAL) | H2SMV-FPV (TEST) | Native Mean (5 sets) | |
| Public Sapiens2-0.4B [ 20 ] | 57.94 | 55.70 | 69.69 | 54.84 | 40.90 | 55.81 | – |
| S2HAND-Pose-0.4B (Ours) | 64.73 | 70.43 | 74.40 | 70.66 | 49.48 | 65.94 | +10.13 |
| Public Sapiens2-0.8B [ 20 ] | 59.06 | 54.78 | 68.03 | 54.21 | 41.78 | 55.57 | – |
| S2HAND-Pose-0.8B (Ours) | 64.60 | 70.12 | 73.66 | 69.68 | 50.01 | 65.62 | +10.04 |
| Public Sapiens2-1B [ 20 ] | 61.71 | 52.53 | 57.17 | 52.26 | 46.76 | 54.09 | – |
| S2HAND-Pose-1B (Ours) | 69.30 | 73.68 | 75.08 | 73.46 | 54.89 | 69.28 | +15.20 |
| Rank | Adaptation Scope | Fisher Trace | Per Parameter | Trainable Params (M) | HInt (VAL) [email protected] (%) |
|---|---|---|---|---|---|
| 1 | Transformer block 38 | 89.620 | 75.76 | ||
| 2 | Transformer block 40 | – | – | ||
| 3 | Transformer block 39 | – | – | ||
| 4 | Transformer block 36 | – | – | ||
| 5 | Transformer block 4 | – | – | ||
| 6 | Transformer block 37 | – | – |
| Rank | Block | Fisher Trace | Rank | Block | Fisher Trace | Rank | Block | Fisher Trace | Rank | Block | Fisher Trace |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | 18 | 7 | 7 | 13 | 20 | 19 | 3 | ||||
| 2 | 17 | 8 | 13 | 14 | 9 | 20 | 0 | ||||
| 3 | 16 | 9 | 12 | 15 | 11 | 21 | 22 | ||||
| 4 | 15 | 10 | 6 | 16 | 10 | 22 | 1 | ||||
| 5 | 19 | 11 | 5 | 17 | 4 | 23 | 2 | ||||
| 6 | 14 | 12 | 8 | 18 | 21 | 24 | 23 |
| Rank | Block | Fisher Trace | Rank | Block | Fisher Trace | Rank | Block | Fisher Trace | Rank | Block | Fisher Trace |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | 22 | 9 | 15 | 17 | 10 | 25 | 28 | ||||
| 2 | 19 | 10 | 20 | 18 | 8 | 26 | 1 | ||||
| 3 | 23 | 11 | 11 | 19 | 26 | 27 | 4 | ||||
| 4 | 17 | 12 | 14 | 20 | 7 | 28 | 3 | ||||
| 5 | 18 | 13 | 9 | 21 | 27 | 29 | 29 | ||||
| 6 | 24 | 14 | 25 | 22 | 6 | 30 | 2 |
| Rank | Block | Fisher Trace | Rank | Block | Fisher Trace | Rank | Block | Fisher Trace | Rank | Block | Fisher Trace |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | 28 | 11 | 23 | 21 | 17 | 31 | 4 | ||||
| 2 | 29 | 12 | 22 | 22 | 10 | 32 | 6 | ||||
| 3 | 30 | 13 | 33 | 23 | 20 | 33 | 39 | ||||
| 4 | 31 | 14 | 21 | 24 | 11 | 34 | 2 | ||||
| 5 | 27 | 15 | 13 | 25 | 19 | 35 | 3 | ||||
| 6 | 26 | 16 | 12 | 26 | 0 | 36 | 1 |
| Rank | Block | Fisher Trace | Rank | Block | Fisher Trace | Rank | Block | Fisher Trace | Rank | Block | Fisher Trace |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | 38 | 15 | 43 | 29 | 19 | 43 | 47 | ||||
| 2 | 40 | 16 | 29 | 30 | 18 | 44 | 10 | ||||
| 3 | 39 | 17 | 30 | 31 | 17 | 45 | 46 | ||||
| 4 | 36 | 18 | 27 | 32 | 1 | 46 | 7 | ||||
| 5 | 4 | 19 | 26 | 33 | 14 | 47 | 8 | ||||
| 6 | 37 | 20 | 28 | 34 | 5 | 48 | 9 |
| Seed | HInt (VAL) | HInt (TEST) | COCO (VAL) | FreiHAND (EVAL) |
|---|---|---|---|---|
| 42 | 68.451 | 67.491 | 47.452 | 61.761 |
| 43 | 68.465 | 67.464 | 47.412 | 61.760 |
| 44 | 68.350 | 67.384 | 47.336 | 61.740 |
| Mean SD |
| System | L AP | L AR | R AP | R AR | Hand AP | Hand AR | Body AP | Body AR |
|---|---|---|---|---|---|---|---|---|
| Public Sapiens2-5B | 53.88 | 58.67 | 49.96 | 55.68 | 51.92 | 57.18 | 79.09 | 82.83 |
| S2HAND-MT (Ours) | 56.56 | 61.96 | 52.95 | 58.86 | 54.75 | 60.41 | — | — |
| S2HAND-Pose (Ours) | 66.11 | 71.25 | 63.51 | 69.51 | 64.81 | 70.38 | 79.28 | 83.07 |
| Head | Metric | Value |
|---|---|---|
| Segmentation | all-pixel agreement | 99.9612% |
| foreground agreement | 99.5956% | |
| hand-mask IoU | 99.3246% | |
| Normal | mean angular drift | |
| mean image-P95 angular drift | ||
| foreground fraction within | 99.9755% |
| COCO Body+Foot AP (%) | Face 16-Anchor NME | Dense Drift | |||||
|---|---|---|---|---|---|---|---|
| Scale | Public | S2HAND-Pose | Public | S2HAND-Pose | 238 Points | ||
| 0.4B | 78.98 | 77.36 | 0.1515 | 0.1252 | 0.2082 | ||
| 0.8B | 79.03 | 78.05 | 0.1550 | 0.1325 | 0.1057 | ||
| 1B | 79.23 | 79.34 | 0.1297 | 0.1324 | 0.1108 | ||
| 5B | 79.09 | 79.28 | 0.1428 | 0.1510 | 0.1132 | ||
| Dataset | System | [email protected] | [email protected] | [email protected] | Tail .20 |
|---|---|---|---|---|---|
| H2O TEST | S2HAND-MT | 70.79 | 53.39 | 87.39 | 3.63 |
| + TailGuard | 71.05 | 53.55 | 87.70 | 3.33 | |
| NewDays TEST | S2HAND-MT | 62.57 | 49.42 | 72.48 | 13.45 |
| + TailGuard | 64.55 | 51.07 | 74.96 | 10.43 |
| Context | Policy | [email protected] | Tail .20 | Trigger (%) | False Re-Crop |
|---|---|---|---|---|---|
| 1.5 | native | 58.54 | 16.96 | 0.00 | 0.00 |
| 1.5 | always if detected | 67.64 | 3.36 | 95.04 | 0.51 |
| 1.5 | per-hand AUC oracle | 68.72 | 3.04 | 67.04 | 0.00 |
| 1.5 | conf. 0.350000 | 61.63 | 12.64 | 16.16 | 0.00 |
| 1.5 | conf. 0.400000 | 62.66 | 11.04 | 22.08 | 0.00 |
| 1.5 | conf. 0.450000 | 63.34 | 9.92 | 28.96 | 0.55 |
| Dataset | System | [email protected] | [email protected] | [email protected] | [email protected] | Trigger (%) |
|---|---|---|---|---|---|---|
| NewDays TEST [ 30 ] | S2HAND-Pose (Ours) | 70.04 | 59.83 | 81.14 | 88.69 | 0.00 |
| + TailGuard | 70.69 | 60.41 | 82.01 | 89.48 | 9.29 | |
| H2O TEST [ 22 ] | S2HAND-Pose (Ours) | 73.96 | 58.08 | 90.79 | 97.62 | 0.00 |
| + TailGuard | 74.23 | 58.30 | 91.11 | 97.97 | 0.39 | |
| H2SMV-FPV TEST | S2HAND-Pose (Ours) | 56.14 | 40.59 | 65.79 | 76.61 | 0.00 |
| + TailGuard | 56.57 | 40.91 | 66.33 | 77.18 | 2.33 |
| Stage | Mean (ms) | Median (ms) | P95 (ms) | P99 (ms) | Max (ms) | N |
|---|---|---|---|---|---|---|
| Native S2HAND-MT | 737.76 | 892.52 | 933.45 | 1804.77 | 6833.52 | 1000 |
| WiLoR-YOLO | 16.21 | 10.15 | 10.62 | 11.76 | 5478.96 | 1000 |
| Conditional S2HAND-MT re-crop | 221.24 | 0.00 | 487.23 | 489.62 | 492.13 | 1000 |
| S2HAND-MT + TailGuard | 975.21 | 963.38 | 1406.02 | 1815.54 | 12312.48 | 1000 |
| TailGuard-triggered samples only | 1246.91 | 1391.30 | 1408.99 | 1775.58 | 2305.06 | 460 |
| Dataset | Hands | Clusters | Public Pose [95% CI] | Hand5 Pose [95% CI] |
|---|---|---|---|---|
| FreiHAND EVAL | 3960 | 3960 | ||
| H2O-4096 | 4096 | 46 | ||
| DexYCB-997 | 997 | 197 | ||
| AssemblyHands-8k | 7982 | 293 | ||
| H2SMV-FPV | 9379 | 306 |
| Hand XY Source | Instance Top-1 (%) | Macro Top-1 (%) |
|---|---|---|
| Uni-Sign native RTMPose | 71.43 | 76.72 |
| Public Sapiens2-5B | 71.43 | 75.98 |
| S2HAND-MT (Ours) | 69.05 | 73.77 |
| S2HAND-Pose (Ours) | 71.43 | 76.72 |
Human Pose Estimation
Human pose estimation recovers people’s body keypoints, 3D skeletons, or meshes from images, video, and other sensor data.
See all 89 papers