- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper proposes CycloFormer, a Transformer that estimates hand pose from a 16-electrode wrist sEMG band and is exactly invariant to cyclic shifts of the electrodes, i.e. to rotating the band. It reports the best landmark distance on all three emg2pose splits (Table 1). It then compares sEMG with monocular vision (HaMeR, HandOccNet) at different levels of fingertip occlusion, on rendered emg2pose poses and on real paired EgoEMG data, and finds that sEMG is more accurate when most fingertips are hidden. Finally, it fits a scaling law over a 32-run grid of model width and training fraction and concludes that the task is data-limited. No new data are collected; the data axis subsamples the existing emg2pose training set.
Critical strengths
The rotation-invariant design is simple and, as far as I can see from App. B.1.1 and the released code, correct: the error is flat over all 16 shifts (Fig. 5). It removes the need for band-rotation calibration, which matters for a device that people put on every day, and the LD gains over vemg2pose (1.1/1.9/1.6 mm, Table 1) are clear. Code and checkpoints are released. The cross-modal setup is also built with care (renderer check on FreiHAND, retargeting floor, real paired data).
Critical weaknesses
- Misleading title and an unsupported scaling claim.
The motivation and design of the scaling study do not hold up, starting with the title. "Scaling Wrist sEMG" suggests that sEMG data are scaled up, but no new data are collected: the study trains eight model widths (1.49M to 84.77M) on 20–100% subsets of the existing emg2pose training set. In my view, the title is therefore kinda misleading and overclaims what the paper actually does. The message that large-scale data collection is a key direction is therefore an extrapolation beyond the largest dataset tested. For contrast, EgoScale (Zheng et al., 2026), which the paper itself cites, trains on 20,854 hours of egocentric human video and reports a log-linear relation between data scale and validation loss. I think this is the right kind of design for testing a scaling law, because the data themselves are scaled up. I also do not see why larger models should be the goal for wrist sEMG, whose appeal is accurate pose from a cheap signal with lightweight, real-time inference on the device. The paper reports neither on-device cost (only A100 time, Table 18) nor a causal model. The fitted law itself is confounded by the training protocol. All runs use 100 epochs and the same learning-rate schedule while the per-GPU batch grows with width (64 to 768, Table 7), so at a fixed GPU count, large models on little data get far fewer updates. This alone could produce the kind of N/D penalty that Eq. 1 adds. The U-shape rests mostly on one single-seed run (84.77M at 20%), and App. E.1 notes that ε∞ is poorly identified, so the claim that the task is data-limited (β > α) is not established. This is the paper's main framing, and fixing it needs new data or a re-run of the grid, which is beyond a rebuttal.
- Confounded comparison with vision.
The occlusion crossover is confounded with hand pose. The paper sorts frames into groups by how many fingertips are occluded (0 to 5). sEMG cannot see occlusion, yet CycloFormer's own error drops in the high-occlusion groups (EgoEMG: 17.30 to 14.56 mm, Table 3; renders: 12.36 to 10.59 mm, Table 25). These groups probably hold different, more closed poses, on which CycloFormer's PA-MPJPE is lower anyway: 11.02 mm on closed hands against 12.71 mm on open hands, and Table 26 reports no open-hand value at 5/5. On real EgoEMG data, vision is more accurate from 0 to 4 occluded fingertips, and sEMG wins only when all five are occluded (14.56 vs. 20.61 mm, Table 3). The paper does not say how many frames, recordings or subjects fall into this 5/5 group. Fully occluded frames are likely rare (the paper notes that most emg2pose frames have few occluded fingertips), so if the group is small or comes from a few subjects, the advantage may reflect chance or a few individuals rather than a general effect. The 95% intervals in Fig. 3b do not settle this, because the paper does not say how they were computed for EgoEMG, unlike the session bootstrap it uses for the renders (p. 8). If frames rather than subjects were resampled, the intervals would look tighter than they should, since frames from the same recording are strongly correlated. The vision side is also not the strongest available: HaMeR (CVPR 2024) is called "the leading vision baseline", but newer models report better accuracy on standard benchmarks, for example WiLoR (CVPR 2025), which the paper itself cites. As it stands, the evidence does not separate "sEMG is robust to occlusion" from "high-occlusion frames are easier poses".
Taken together, the framing in the title and abstract and two of the three contributions are not supported by the current evidence. This is why I recommend rejection despite the solid invariance result.
Questions for authors
Positive answers to Q2 and Q3 would move me toward weak rejection; the title and the lack of new data would remain. Q1 and Q4 affect how far I can trust the reported gains.
- Attention implementation (clarification).
App. B.1 states that applying the rotary embedding before W_q and W_k "destroys the symmetry". Yet the released code keeps "legacy" attention modes for "pre-RoPE-bug-fix checkpoints" (cs_tds_ct_arch.py, l. 129–152): one applies RoPE before the projections, and two rotate over the head axis instead of the electrode or time axis, so they carry no positional information. Were all 80× results (e.g. Table 2, Fig. 5, Table 27) produced with rope_mode="fixed"? If any were produced with a legacy mode, those results, in particular the CRoPE ablation in Table 2, would need to be redone.
- The EgoEMG 5/5 group (clarification and analysis).
How many frames, recordings and subjects fall into each EgoEMG occlusion group, especially 5/5? Were the 95% intervals in Fig. 3b computed by resampling frames or subjects? Please also check whether the 5/5 advantage holds under two stricter tests. (a) A subject-level bootstrap, which resamples whole subjects rather than single frames. Frames from the same subject and recording are highly correlated, so resampling frames makes the intervals too narrow, and an advantage that comes from only a few subjects would not survive this test. (b) A pose-matched comparison, which compares sEMG and vision on the same or similar hand poses at different occlusion levels, for example by rendering each emg2pose pose from several viewpoints, or by comparing within groups of similar poses on EgoEMG. The high-occlusion groups are likely dominated by closed hands, on which CycloFormer is more accurate anyway (11.02 vs. 12.71 mm, Table 26), so this test separates the effect of occlusion from the effect of hand pose.
- The scaling grid (additional analysis).
Does the upturn at (84.77M, 20%) persist when that run gets as many updates as the smaller models (or uses the same batch size) and a tuned learning rate? Please also report the training loss for each cell (at present it is shown only for the 1.49M and 4.18M seed runs, Fig. 6).
- Causality and deployment (clarification and analysis).
The temporal attention has no causal mask and the convolutions use symmetric padding (cs_tds_ct_arch.py, e.g. l. 511 and l. 779), so each frame also sees future samples within the 5-s window. vemg2pose, in contrast, uses a left context and a sequential LSTM rollout (App. B.4.4–B.4.5). Are the Table 1 comparisons made under the same causal constraint, and what is the latency on wearable-class hardware?