acceptodds
Under review as a conference paper at ICLR 2027

ATLAS: Timed Lexical Anchors for Sign Language Production at Scale

Abstract

Glosses, ordered transcriptions of signs as written words, considerably improve sign language production (SLP) when used as a condition, but require expert annotation that large signing corpora lack. Glosses derived from the text alone do not reproduce this gain, as they carry no information about when each sign occurs. We instead extract timed lexical anchors from the signing itself with a sign-identity model trained on isolated sign dictionaries, and combine them with the sentence as a hybrid condition: the anchors supply the timing of the signs they cover and the sentence supplies the remaining content. At inference, the anchors are predicted from the sentence alone. The scheme requires no gloss annotation of the training corpus and therefore extends to large unannotated datasets. We extract and release SMPL-X parameters for more than 1,700 hours of video from YouTube-SL-25, and train ATLAS, a continuous-latent flow-matching generator, on more than 600 hours of ASL signing, substantially more than used in prior work. The anchors improve back-translation at every data scale we test: an anchored model trained on a fraction of the data surpasses a sentence-only model trained on all of it, and relative to each system’s ceiling, ATLAS reaches more than twice the back-translation score of the strongest baseline. Finally, we audit the SLP evaluation protocol used in the literature with content-free baselines, outputs that carry no signing by construction. Reference-based distance metrics favour low-amplitude motion, to the extent that such outputs can score higher than state-of-the-art models. We propose a protocol that reports the model’s performance grounded in these baselines and measures content with a public back-translation model.

Primary area:
generative models
Keywords:
sign language production, human motion generation, flow matching

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_MXdyOpenReview ↗
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
  1. 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.

  2. 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
  1. 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 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.

27k generator trained withBLEU-4, mean ± stdvs sentence-only [95% CI]
sentence only1.12 ± 0.07
caption keywords in caption order1.17 ± 0.05+0.05 [−0.17, +0.27]
spotted signs moved to caption positions1.18 ± 0.15+0.07 [−0.13, +0.27]
caption force-aligned to the motion, every keyword placed1.28 ± 0.11+0.16 [−0.01, +0.32]
ATLAS: spotted anchors at their signed positions1.72 ± 0.08+0.59 [+0.39, +0.79], p < 0.001

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.

Training sentencesSentence-onlyATLASGain
27k (How2Sign only)0.490.95+0.47
100k0.391.62+1.23
404k0.412.10+1.69

All three gains have p < 0.001.

  • The gain is larger where the pairing is weaker: +1.69 against +0.77 on How2Sign for the 404k model.
  • Only the anchored model benefits from more data. The sentence alone does not tell the generator which signs to produce, and more data does not change that: the sentence-only model stays at about 0.5, while ATLAS more than doubles from 27k to 404k.
  • It holds where signing departs from the caption. On clips whose spotted order differs from caption order the gain is +1.76 [+1.13, +2.37].

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.

Reviewer_4vHVOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ATLAS, a sign language production framework that uses timed lexical anchors to guide motion generation. The authors extract lexical anchors and their temporal positions from continuous signing using isolated-sign dictionaries, and train a predictor to infer these anchors from text at inference. The predicted anchors are combined with sentence embeddings to condition a flow-matching generator. The paper also releases large-scale SMPL-X motion data from YouTube-SL-25 and examines the limitations of existing SLP evaluation metrics. Experiments on How2Sign show improvements over the compared methods.

Critical strengths
  1. The idea of predicting lexical anchors and their temporal positions from input sentences is interesting and well motivated. It offers a reasonable way to improve SLP without requiring gloss annotations for the target training corpus.
  2. The authors use isolated-sign dictionaries to extract lexical and temporal information from continuous signing videos, allowing the model to leverage large-scale datasets without additional gloss annotations. The experiments show consistent improvements with the proposed anchors across different training data scales.
Critical weaknesses
  1. The gap between isolated and continuous signing may affect lexical spotting accuracy. Although the authors introduce several adaptations, the accuracy of the extracted anchors and their temporal boundaries is not directly evaluated against ground-truth annotations.
  2. While Tables 2 and 4 show the benefits of temporal information, they do not clearly distinguish the contributions of gloss ordering, temporal positioning, and duration prediction. It is also unclear how the predicted sign durations are preserved when the anchor sequence is uniformly mapped to motion latents.
  3. The discussion of recent SLP methods is insufficient. The authors should consider including the following relevant works:

[1] FGDM: Focal–General Diffusion Model with Semantic Consistent Guidance for Sign Language Production. CVPR2026

[2] SignPR: A Progressive Vector-Quantized Diffusion Framework for Sign Language Production. CVPR2026

[3] Hybrid Autoregressive-Diffusion Model for Real-Time Sign Language Production. ACL2026

Questions for authors

See Weaknesses.

3 comments

Thank you for the positive assessment and the precise questions.

Direct anchor evaluation. You are right, the submission does not evaluate the anchors directly against GT as How2Sign and YouTube-SL-25 do not provide timed gloss annotations. We ran this evaluation on NCSLGR, a public corpus of continuous ASL which contains glosses with start and end timings. It was not used for training any component of ATLAS. We used the 20 narrative videos from the dataset for testing (about one hour of video) which contain 1,001 translated utterances and 7,746 gloss tokens. 38% of the tokens are pointing, classifiers, non-lexical gestures or fingerspelling, which our dictionary does not cover. The corpus also labels some signs differently to our vocabulary (e.g. GIRL+MARRY for WIFE, ns-BOSTON for BOSTON). Therefore, we first create an equivalence table that maps corpus labels to our dictionary labels. It covers label ambiguity, compounds written out, inflected and multi-word forms, notation and numeral variants, and name signs. Label differences not covered by the table count as errors, so the evaluation figures are lower bounds.

Our spotter proposes a candidate sign for each caption word that has a dictionary entry, and accepts a candidate only where it finds a matching span in the motion. For correct anchors we also measure the distance between the centre of the anchor and the centre of the annotated sign. Three controls separate what checking the motion adds from what the caption alone provides:

  • Mismatched caption. The spotter runs on the caption of a different video, so every candidate is wrong. This is the chance level.
  • No motion check. Every candidate is accepted without looking at the motion. This is what the spotter would score if the motion check carried no information.
  • Text-only timing. Each correct sign is placed where its word falls in the sentence. This is the timing that the text can supply.

The results are provided in the tables below:

Anchors source# Per clipCorrect sign
mismatched caption1.13%
no motion check5.845%
spotter2.268%
Correct anchorsWithin 0.25 s of the signMedian error
text-only timing34%0.39 s
spotter71%0.13 s

The evaluation on NCSLGR shows that recognition learnt from isolated dictionary signs carries over to continuous signing. Checking the motion raises the share of correct anchors from 45%, the level of the caption alone, to 68%, and the spotter's score tracks correctness (53% correct in its least confident quarter of anchors, 84% in its most confident). The spotter's timing provides an even larger gain: a correct anchor lies within 0.25 s of the annotated sign in 71% of cases, against 34% when the sign is placed by its word's position in the sentence, and this timing is what the text alone cannot supply.

Order, position and duration. In short, at inference the generator relies on which signs are named and on their order; the signed positions matter in training, where they make the condition learnable. To isolate these factors, we kept the trained ATLAS generator fixed and, at inference, changed one property of the predicted anchors at a time: which signs are selected, their order, their positions, their durations. The results (paper's sampling seed) are provided in the following table:

Predicted anchorsHow2Sign 27kHow2Sign 404kYouTube-SL-25 404k
none (sentence-only model)1.191.480.41
sign order shuffled, time slots kept1.511.791.03
order kept, evenly spaced1.721.972.11
every duration set to 0.5 s1.722.202.02
unchanged (ATLAS)1.782.252.10
  • Which signs are named matters first: even shuffled, the anchors beat the sentence-only model (+0.32, +0.31, +0.62; p ≤ 0.03).
  • Order adds +0.26, +0.47 and +1.07 (p ≤ 0.006). The shuffled signs are still produced, in the shuffled order (92% are found by the sign-identity model), which the back-translation and BLEU's n-grams penalise.
  • Exact position matters little at inference once the order is kept: spacing the anchors evenly lowers back-translation by at most 0.28 (significant only on How2Sign at 404k). Position matters more in training (below).
  • Duration has no measurable effect: setting every duration to 0.5 s changes BLEU-4 by less than 0.1. Possible reasons for this are discussed below.

We also trained the 27k generator with the same spotted signs placed at different positions, and gave every anchored model the same anchors at inference, spotted in the reference motion (hence 1.62 for ATLAS rather than 1.78 with predicted anchors). Only the timing seen in training differs.

27k generator trained withBLEU-4
sentence only1.19
spotted signs moved to caption positions1.03
spotted signs at scrambled positions1.07
spotted signs at their signed positions (ATLAS)1.62

ATLAS is above both controls (+0.59 and +0.55, p < 0.001), and neither control meaningfully improves on the sentence-only model. To learn the condition, the generator must be trained on anchors placed where their signs actually occur in the motion; once trained, it tolerates imprecise placement. The other training controls are in our reply to Reviewer MXdy. We added both tables to the ablation section of the paper, separating the changes made at inference from the retrained controls.

How durations survive uniform tiling. Each anchor becomes one token, and each stretch without an anchor becomes one <unk> token per 0.5 s. The tokens are then spread uniformly over the latent steps. Durations are therefore encoded indirectly by design: a sign's span sets where the next gap begins, and so how many <unk> tokens separate it from the following sign. The condition encodes the order of the signs and, through the gaps, their timing. After tiling, an anchor's token sits a median 0.11 s from the centre of its span. A sign's length is not a separate input, and it does not need to be as spotted signs last 0.39 s on average (80% between 0.33 and 0.50 s), close to the width of one token. A generator conditioned on exact spans does no better (1.88 vs 1.78, n.s.). We added this explanation to Section 3.3 of the paper.

We also extended the Related Work section of the paper to include the recent SLP approaches you indicated.

Thanks for the response. I have no further questions and will keep my score unchanged.

Reviewer_fLqwOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on sign language production. The authors point to the fact that gloss condition does help sign language production , but it's costly as it requires manual human annotations. Also, the glosses obtaiend from the text too arent performing as well for many reasons. So they leverag sign language dictiionairy to create (weak) time aligned glosses in for sign language videos, and shows this with backtranslation sets good performances. Also, the authors are releasing pre-computed features.

Critical strengths
  1. The anchor processing to get glosses is clean. Levarage on sign dictionary datasets to have some sort of time aligned pseudo glosses .
  2. Hybrid conditioning. still the text provides the rest of the information the anchors couldnt get.
  3. Better performance (I wouldnt say strong see the weakenss) compared to other models
  4. Releases features in the future.
  5. Leverages naturual signing datasets as YouTube-SL-25 which I really think the community should start using more these kind of datasets than just old interpreted signing datasets.
Critical weaknesses
  1. I think though the BLEU-4 improved by rough 1 to 2 points, I think at this very low scores , it's basically mostly largely unsignificant.
  2. Adding other metrics as BLEURT might be good.
Questions for authors
  1. Can you additionally test your method on a few samples of naturally signing from YouTube-SL-25 that wasnt used in training
  2. Can you also provide BLEURT scores.
2 comments

Thank you for the encouraging assessment.

In short: the gains are significant (p < 0.001 in every sampling seed), BLEURT, chrF and ROUGE-L metrics agree, and on unseen natural signing from YouTube-SL-25 the gain is larger (+1.69 BLEU-4).

Are 1–2 BLEU points meaningful at such low scores? The gains may seem small in absolute terms but they are large relative to the range that this metric resolves. Through the same translator, outputs with no signing at all score 0.5–0.75 BLEU-4, and the reference GT motion scores 10.6. SOKE and SignSparK (0.68 and 0.67) are near the floor level. ATLAS reaches 2.25, three times the still pose and 52% above the sentence-only model (1.48). Relative to the ceiling of each fitting, this is 21% for ATLAS, 14% for the sentence-only model and 9% for SOKE and SignSparK.

We run a paired bootstrap over the test clips (10,000 resamples) and two more sampling seeds to check the robustness of the numbers. The gain stays significant in every seed (p < 0.001), and a single score moves by only 0.1–0.2 BLEU-4 between seeds.

ATLAS vs sentence-onlyΔBLEU-4pMean ± std over 3 sampling seeds
How2Sign, 404k training sentences+0.77< 0.001+0.71 ± 0.14
How2Sign, 27k+0.58< 0.001+0.60 ± 0.07
YouTube-SL-25, 404k+1.69< 0.001+1.44 ± 0.22

The gain also recurs in every independently trained pair of generators in the scaling experiment, with 27k, 100k and 404k training sentences (Table 5).

Additional metrics. We add three metrics, computed on the same back-translations as BLEU-4: BLEURT-20, a learned measure of semantic similarity (the checkpoint used by YouTube-ASL and Uni-Sign); chrF, which gives credit to partly correct words through character n-grams; and ROUGE-L, which rewards the longest in-order word overlap. On the How2Sign test clips, all three agree with BLEU-4 and rank ATLAS first among the generated outputs:

SystemBLEU-4chrFROUGE-LBLEURT
reference motion10.6229.729.90.445
ATLAS (404k)2.2516.612.60.337
sentence-only (404k)1.4814.610.10.312
SignSparK / SOKE0.67 / 0.6813.5 / 11.58.4 / 6.90.298 / 0.281
still pose0.7511.68.00.313
unrelated sentence0.4913.57.80.282

BLEURT has a narrow range here: it gives partial credit to any fluent, non-committal sentence. The translator answers a still pose with exactly such sentences, so the still pose already scores 0.313. Correct signing of an unrelated sentence scores lower, 0.282, because its content is specific but wrong. ATLAS is the only scored system above the still-pose row in all metrics. Its BLEURT gain over the sentence-only model is +0.025 at 404k and +0.018 at 27k, p < 0.001.

YouTube-SL-25. As you suggested, we tested ATLAS on natural signing from YouTube-SL-25 that was not used in training: 2,000 ASL clips from 465 videos of channels held out from the training of every ATLAS component. The public Uni-Sign model scores only 0.9 BLEU-4 on the reference motion of these clips, so we fine-tuned it on our YouTube-SL-25 ASL training split. We will release the checkpoint for reproducibility. The fine-tuned model scores 10.25 on the test references and provides the following results for ATLAS:

YouTube-SL-25 testBLEU-4chrFROUGE-LBLEURT
reference motion10.2529.029.10.418
ATLAS (404k)2.1018.315.70.315
sentence-only (404k)0.4113.88.70.256
still pose0.407.54.50.214

The anchors improve every metric compared to the sentence-only model, and by more than on How2Sign: +1.69 BLEU-4 and +0.059 BLEURT (both p < 0.001). On natural signing the sentence-only model stays at about 0.5 BLEU-4 at every training size, while ATLAS grows from 0.95 (27k) to 2.10 (404k). Additional details on model scaling on YouTube-SL-25 are provided in our reply to Reviewer MXdy.

We also added several qualitative examples of the YouTube-SL-25 outputs to the supplementary material.

My concerns have been addressed. I maintain my postive score.

Comments on the paper

1 comment

We thank the reviewers for their careful and constructive reviews. We are glad that all three found the core idea, timed lexical supervision without gloss annotation, a well-motivated way to scale lexical conditioning to large unglossed corpora, and that the hybrid conditioning, practical evaluation protocol and YouTube-SL-25 data release were seen as useful contributions. The individual replies answer each question with new experiments and clarifications. In summary:

  1. Additional testing (MXdy, fLqw). On 2,000 YouTube-SL-25 ASL clips from channels held out from training, where sentence and signing are not paired by construction, ATLAS raises BLEU-4 from 0.41 to 2.10 (+1.69, p < 0.001), more than on How2Sign (+0.77). Only the anchored model improves with more training data: from 0.95 at 27k to 2.10 at 404k sentences, while the sentence-only model stays at about 0.5 BLEU-4.
  2. How2Sign pairing (MXdy). To test whether the gains come from How2Sign's construction, we trained control generators whose lexical conditions are derived from the caption instead of verified in the signing. They stay close to the sentence-only model (1.17–1.28 vs 1.12 BLEU-4), while ATLAS reaches 1.72 (p < 0.001, three sampling seeds). Conditions derived from the caption alone cannot be learned from, the anchors must be verified in the signing.
  3. Significance and additional metrics (fLqw). The gains over the sentence-only model, on How2Sign (27k and 404k) and on YouTube-SL-25, are significant under a paired bootstrap over test clips and in each of three sampling seeds (p < 0.001). Additionally, BLEURT, chrF and ROUGE-L agree with BLEU-4 on both test sets. On How2Sign, content-free outputs score 0.5–0.75 BLEU-4 and the reference motion scores 10.6; ATLAS reaches three times this floor, while SOKE and SignSparK stay near it.
  4. Anchor accuracy (4vHV). On NCSLGR, a continuous ASL corpus with timed glosses that no ATLAS component was trained on, 68% of the spotted anchors name a sign annotated in the utterance (45% without checking the motion, 3% at chance), and 71% of the correct anchors lie within 0.25 s of the annotated sign (34% with timing from the text).
  5. Contributions of order, position and duration (4vHV). At inference, the gain rests first on which signs are named and then on their order. The trained generator tolerates imprecise timing: evenly spaced anchors or a fixed duration of 0.5 s keep most of the gain, which matters because the anchors at inference are predicted from text. However, timing is essential in training. The same signs at caption or scrambled positions bring no gain (1.03 and 1.07 vs 1.62), so the anchors must be placed where they are actually signed.
  6. Positioning against prior work (MXdy). SIGNER and NAT-EA need a complete gloss transcript of the target corpus. ATLAS's spotter instead verifies each caption candidate in the motion, does not assume caption order, and uses an encoder trained on isolated dictionary signs. This is what allows training on large unglossed corpora. A control that keeps NAT-EA's monotonic alignment of the full caption trains a generator to only 1.28 against 1.72 BLEU-4 (item 2).

Changes to the paper. We added SIGNER, NAT-EA, FGDM, SignPR and HybridSign to the Related Work section and credited NAT-EA's mean-pose and random-pose baselines as a precedent for our proposed evaluation protocol. Section 3.3 now explains how anchor durations are encoded. The ablation section now separates the changes made at inference from the retrained generators and includes the training controls. The supplementary material now contains new qualitative examples of generated YouTube-SL-25 clips.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.