- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
ATLAS is a flow-matching SMPL-X sign generator conditioned on timed lexical anchors that a dictionary-grounded sign-identity model spots in the signing.
Critical strengths
-
Annotation-free timed lexical supervision. Spotting yields (sign, span) pairs on 377k unannotated clips, precisely the bottleneck that confines lexical conditioning to small glossed corpora. This is the paper's only defensible delta, and it is a real one.
-
A practical reporting protocol. Content-free baselines, per-frame and accumulated DTW, duration and amplitude, and scores as a fraction of the ceiling would make SLP tables far more comparable.
Critical weaknesses
The benchmark may already supply the alignment the method claims to recover. Every How2Sign clip renders a single written English sentence. The signer watches that sentence as subtitles and signs it while the source is played at 0.75 of normal speed. Sentence and clip are therefore paired one to one by construction, and the clip's duration follows the source sentence rather than the signing. On such data, a generator that puts the sentence's content words in sentence order and takes its output length from the sentence is already close to the right answer, and that is what the anchor predictor and the length head do. A 50% gain from predicted anchors therefore does not show that anchors help where sentence and signing genuinely disagree. Nor is that setting evaluated as YouTube-SL-25 supplies 377k training clips and no test result.
Questions for authors
- The core mechanism, lexical sequence with text-predicted durations, appears in SIGNER (ECCV 2026, arXiv:2506.07460; previously know as GLOS when not accepted but now SIGNER), which is not cited. It is acceptable to treat SIGNER as a concurrent work. But would you please clarify what ATLAS adds beyond replacing annotated glosses with spotted anchors? Beyond SIGNER, how is the anchoring scheme positioned against NAT-EA, and how significant is your modified part?
3 comments
Thank you for pointing us towards SIGNER and NAT-EA. We modified the Related Work section of the paper to include them.
Conceptually, SIGNER, NAT-EA and ATLAS share an intermediate representation, an ordered sequence of lexical units expanded over time, but differ in where the units and their timing come from. SIGNER converts input text sentences into full gloss sequences with an off-the-shelf text-to-gloss model, predicts a duration for each gloss and repeats each embedding for that duration. It is trained and evaluated on PHOENIX14T and CSL-Daily, which are glossed, and its authors note that a text-to-gloss model must be trained for each new dataset. NAT-EA is a gloss-to-pose model trained on PHOENIX14T whose aligner places the known GT gloss sequence monotonically on the poses. Both rely on a complete annotated gloss transcript of the target corpus, which typically only comes with small lab-recorded datasets. How2Sign and YouTube-SL-25 have none, so neither method can be trained on them directly.
What ATLAS adds beyond SIGNER. Replacing annotated glosses with spotted anchors is what unlocks training on large unglossed corpora, and the anchored model benefits from the scale: tested on unseen YouTube-SL-25 videos, ATLAS improves from 0.95 BLEU-4 when trained on the 27k How2Sign sentences to 2.10 with 404k, while the sentence-only model stays at about 0.5 (previous comment). This replacement is not a drop-in substitute. The annotation-free version of SIGNER's condition, caption keywords in caption order, gives no significant gain over the sentence-only model (1.17 vs 1.12, previous comment), and on CSL-Daily, where GT glosses exist, caption words at text-derived positions leave gloss WER at 0.90 (sentence only: 0.92) while the same words at the true positions of their signs give 0.75 (Table 2 of the paper). The lexical units help the most when placed where they are actually signed, and since spotted anchors are sparse, ATLAS keeps the sentence in the condition and marks the unanchored stretches explicitly. SIGNER's local temporal fusion is orthogonal and could replace our uniform tiling.
Positioning against NAT-EA. ATLAS's spotter plays the role of NAT-EA's external aligner: it places lexical units on the training motion, and the generator learns from the timed units. Built for corpora without a gloss transcript, it differs in three ways, each of which matters downstream. 1) It detects rather than aligns: it accepts only the caption candidates it finds in the motion, with a threshold calibrated on captions of other videos so that false anchors are bounded without labels; back-translation peaks at this threshold and falls for stricter and looser ones (Table 4). 2) It does not assume caption order: each candidate is located separately, and the same spotted signs moved to their caption positions train a generator to only 1.18 BLEU-4, against 1.72 at their signed positions (previous comment). 3) Its encoder is trained on isolated dictionary signs rather than on annotations of the target corpus, and its design (span-level training, a time-reversal term, dictionary exemplars, self-training) raises the number of correct anchors by 40% at the same false-anchor budget (Table 3). Keeping NAT-EA's assumptions with our encoder (complete caption, monotonic, every word placed) trains a generator to 1.28 BLEU-4 against 1.72 for ATLAS (p < 0.001), not significantly above the sentence-only model (previous comment). Against GT glosses on continuous ASL, checking the motion raises the share of correct anchors from 45% to 68% (our reply to Reviewer 4vHV).
We also credited NAT-EA's mean-pose and random-pose baselines as a precedent for our content-free rows.
Thank you, most of my concerns are addressed. Icannot give clear acceptance, as the novelty is curtailed due to related work, but give my best wishes.
Thank you for the thoughtful comments and analysis of the How2Sign capturing protocol.
We tested whether the alignment of the text and sign sequences explains the ATLAS gains in two ways: 1) on How2Sign itself, by training generators on several control conditions, and 2) on the unseen test set of YouTube-SL-25 where sentence and clip are not paired by construction.
How2Sign. If the pairing already places caption words where they are signed, a generator trained with caption keywords should match ATLAS. We trained the 27k generator (the model that sees only the 27k How2Sign training sentences) on exactly the condition you describe (every caption keyword, in caption order, length inferred from the sentence) and on two related conditions, the results are shown in the table below.
To separate real gains from sampling noise, we ran the inference of every model under three sampling seeds. The reported BLEU-4 scores are mean ± std over the seeds. Each row is compared with the sentence-only model by a paired bootstrap over test clips (10,000 resamples), pooled over the seeds; an interval that includes zero means the gain is not significant.
ATLAS outperforms every control (+0.43 to +0.54, p < 0.001). We also checked that the controls utilise their anchors: withholding them costs 0.18–0.44 BLEU-4, against 1.01 for ATLAS. Yet, caption keywords, and spotted signs moved to caption positions, add no significant gain over the sentence alone. Forced alignment, the only control whose timing comes from the motion, gains +0.16, at the edge of significance: timing from the motion helps a little on its own. It reaches only a quarter of ATLAS's gain because it must place every caption word in caption order, including words that were never signed. ATLAS keeps only signs its spotter verifies in the motion, in whatever order they occur. This verification is what makes the annotation-free supervision learnable.
YouTube-SL-25. We also test on YouTube-SL-25, where signing is natural and was captioned afterwards, so the order of the spotted signs follows the caption less than on How2Sign (Kendall τ 0.48 vs 0.61). We partitioned the dataset by source channel: all ATLAS components were trained on 377k sentences from the training channels, and tested on 2,000 ASL clips from 465 held-out videos. The public Uni-Sign model did not work reliably on this domain (0.9 BLEU-4 on reference motion), so we fine-tuned it on reference motion of the training split. The fine-tuned model scores 10.25 BLEU-4 on the test references, we will release the checkpoint for reproducibility.
All three gains have p < 0.001.