{"id":"114a00ce-b3c2-496a-9f64-56430a29b8f4","arxiv_id":"1908.05436","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Forecasting future poses can be done accurately by encoding joint trajectories with a discrete cosine transform and feeding the coefficients to a graph network that learns joint dependencies automatically.","lead":"This paper predicts future human body poses with a feed-forward network that represents each joint's trajectory as a small set of cosine coefficients and learns which joints depend on each other. The method reports the best short- and long-term accuracy on three standard motion-capture benchmarks, especially when trained directly on 3D coordinates.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"3D SOTA claim rests on incomplete baseline comparison and lacks variance estimates","rationale":"The reader's weakest assumption focuses on the DCT smoothness premise. This concern is real but substantially mitigated by the paper itself: for angle trajectories, which are acknowledged to have discontinuities, the authors use the full DCT frequency (20/35 coefficients), and for 3D trajectories the supplementary figures 4 and 7 empirically demonstrate smoothness and show that truncation at 15-30 coefficients retains predictive signal. Thus the smoothness assumption is not fragile for the benchmark results that anchor the central claim. The more load-bearing concern is the external validity of the 3D SOTA claim: the comparison set in Tables 2, 4, and 5 omits the strongest adversarial baseline, AGED, and relies on author-adapted 3D versions of the other baselines. Without variance estimates, one cannot assess whether the margins, while large in per-action cases, are statistically significant at the reported average. The internal inconsistency at 1000 ms in angle space further weakens the unqualified 'state of the art' phrasing. These issues do not invalidate the method's usefulness, but they make the headline claim conditional on a more complete and statistically grounded comparison. The reader already reached a CONDITIONAL verdict; my read reinforces that verdict without moving it, so the recommended verdict is UNCHANGED.","tokens_in":19502,"tokens_out":12854,"duration_ms":123881,"concrete_test":"Re-run Tables 2, 4, and 5 with the released code, replacing the author-adapted 3D baselines with original 3D checkpoints or best-tuned reimplementations of [16] and [17], and include AGED retrained with a position-space loss on 3D coordinates. Compute means and standard deviations over 5 random seeds. If the average MPJPE margin at 400 ms on H3.6M between Ours 3D and the best baseline remains larger than two standard deviations, and if AGED-3D does not beat Ours 3D, the SOTA claim stands; otherwise it must be qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim that the 3D variant 'outperforms the baselines by a large margin' (Section 4.3) is supported only by comparisons against two baselines (Residual sup. [17] and convSeq2Seq [16]) that the authors themselves adapted to 3D coordinates (Section 4.2). The strongest adversarial baseline, AGED [9], is absent from all 3D tables because its code is unavailable. No variance estimates are reported anywhere, so it is unknown whether the reported margins are statistically robust. Additionally, Table 3 shows the angle-space variant is not state-of-the-art at 1000 ms (average 1.27 vs AGED w/adv 1.09), contradicting the unqualified 'state-of-the-art in all cases' statement in the introduction. Thus the strongest claim of 3D superiority is not yet established against the full set of contemporary baselines.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a feed-forward network for human motion prediction that encodes the observed pose sequence in trajectory space via the Discrete Cosine Transform (DCT) and models spatial joint dependencies with graph convolutional layers whose adjacency matrices are learned during training. A residual formulation is used: the network predicts DCT-coefficient residuals relative to the DCT coefficients of the observed sequence padded by replicating the last pose, motivated by the zero-velocity baseline of [17]. The method is evaluated on Human3.6M, CMU-Mocap, and 3DPW, in both joint-angle and 3D coordinate representations, and is reported to outperform prior RNN and convolutional baselines. The paper also argues that angle-space evaluation is unreliable because multiple angle sets can map to the same 3D pose, and that training directly on 3D coordinates gives the best results.","tokens_in":19653,"tokens_out":6570,"duration_ms":62943,"significance":"If the results hold, the paper makes a useful contribution to human motion prediction: it provides a simple feed-forward alternative to RNN-based methods, shows that DCT-based trajectory encoding removes the need to manually set temporal convolutional filter sizes, and demonstrates that learning graph connectivity is preferable to using a fixed kinematic-tree graph. The ablations in Tables 6 and 7 individually justify the DCT encoding, the replicate-padding strategy, residual connections, and the learned connectivity. The supplementary sensitivity analysis of the number of DCT coefficients is informative and honestly reports that angle trajectories are not smooth and therefore require full-frequency DCT, which confines the benefit of coefficient truncation to smooth 3D trajectories; the stress-test concern about non-smooth trajectories is thus already addressed by the manuscript's own analysis.","major_comments":[{"comment":"The abstract states that the proposed approach 'achieves state of the art performance' and Section 1 claims 'state-of-the-art results in all cases.' These blanket claims are contradicted by Table 3: in long-term joint-angle prediction at 1000 ms, the average error of the proposed method is 1.27, which is higher than AGED w/adv's 1.09. While the authors acknowledge this in Section 4.3, the abstract and introduction remain unqualified. Please revise the claims to specify that the state-of-the-art result holds for short-term angle prediction and for 3D-coordinate prediction, with the caveat that the 3D comparison set is limited as described in the next comment.","section":"Abstract and Section 1"},{"comment":"The 3D-coordinate comparisons omit AGED [9], the strongest prior method in angle space, because the authors state that its code and angle predictions are unavailable. Consequently, the statement in Section 4.3 that 'we outperform the baselines by a large margin, particularly when training directly on 3D' is supported only against Residual sup. [17] and convSeq2Seq [16], both adapted by the authors to accept 3D input. This does not establish state-of-the-art 3D prediction against the full field. The authors should either soften the claim to 'outperform the compared baselines' or make a stronger effort to include AGED in 3D (for example, by obtaining or converting AGED's published angle outputs if they become available), or clearly state the comparison scope in the abstract and conclusion.","section":"Section 4.2 and Tables 2–5"}],"minor_comments":[{"comment":"Section 3.4 says training is performed in the exponential map representation of joint angles, while Section 4.2 says the evaluation metric is Euclidean distance in Euler angle representation. Please clarify whether the exponential-map predictions are converted to Euler angles before computing the error, or whether this is a typo.","section":"Section 3.4 and Section 4.2"},{"comment":"No uncertainty estimates (standard deviations, confidence intervals, or repeated runs) are reported. Please state whether the reported numbers are from a single training run or averaged, and consider adding error bars to the key comparisons.","section":"Tables 1–5"},{"comment":"The 'pre-defined connectivity' baseline used in Table 7 is not fully specified; it is unclear whether it uses the kinematic tree with a normalized adjacency matrix, a full graph, or some other construction. Please provide the exact definition so the ablation is reproducible.","section":"Table 7"},{"comment":"The sentence 'to predict the future10 frames' is missing a space; it should read 'future 10 frames.'","section":"Section 4.1"},{"comment":"The checkmark notation in Table 6 is not immediately self-explanatory; a legend or a sentence stating which components are active in each row would improve readability.","section":"Table 6"}],"recommendation":"major_revision","confidential_remarks":"The paper is methodologically sound, well-written, and ships code, but the central state-of-the-art claim is overstated relative to the paper's own Table 3, and the 3D claim is weakened by the absence of AGED. A major revision that scopes the claims accurately and adds a brief note about single-run results would bring it in line with the evidence. The angle-ambiguity observation is interesting and could be expanded, but it is not required for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid empirical paper that deserves a serious referee, but the headline claim is stronger than the tables support. The genuinely new part is combining DCT-based trajectory encoding (borrowed from nonrigid SfM) with a residual feed-forward GCN whose per-layer adjacency matrices are learned. That combination is not in the cited prior work, and the paper shows it works well on H3.6M, CMU-Mocap, and 3DPW.\n\nWhat the paper does well: the ablations in Tables 6 and 7 isolate each component—DCT, padding strategy, residual connections, and learned connectivity—and show each helps. The supplement's analysis of DCT coefficient count is honest, including the finding that angle trajectories have large discontinuities and need full frequencies, while 3D trajectories are smooth enough to truncate. The angle-vs-3D discussion around Figure 4 is a real contribution: it gives a concrete example where angle-space error misleads about 3D pose quality. Code is released, so the results are checkable.\n\nSoft spots, in proportion: the stress-test note is correct. The strong 3D claim is only compared against Residual sup. and convSeq2Seq, both adapted by the authors to 3D. AGED—the strongest angle-space baseline—is absent from the 3D tables because its code is unavailable. The authors disclose this, and it is a reasonable practical constraint, but it means the 'large margin' in 3D is not established against the full contemporary baseline set. Second, Table 3 shows the angle-space variant is not SOTA at 1000 ms (ours 1.27 vs AGED w/adv 1.09), which directly contradicts the abstract and introduction's 'state-of-the-art in all cases'. They do hedge later in Section 4.3, but the framing needs fixing. Third, there are no variance estimates anywhere; given how close some margins are (e.g., on 3DPW), it is unknown which comparisons are statistically robust. That is a common omission in this literature but worth flagging.\n\nThe DCT smoothness assumption is not a fatal flaw—the paper itself shows where it holds and where it does not—but it does mean the core advantage is representation-dependent. Anyone applying the method to a representation without smooth trajectories should expect to keep more coefficients.\n\nMy recommendation: send it to peer review. The experiments are reproducible, the ablations are careful, and the angle-vs-3D argument is worth engaging. The needed revisions are mostly in claim tempering and reporting, not in redoing the work.","headline":"Solid, reproducible motion-prediction paper with a genuinely useful DCT-in-trajectory trick and good ablations, but the 'SOTA in all cases' claim needs tempering because the strongest adversarial baseline is absent from 3D tables.","tokens_in":20216,"tokens_out":2283,"would_cite":true,"duration_ms":26097,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A feed-forward network that predicts residual DCT coefficients in trajectory space, with learned graph connectivity, outperforms recurrent and convolutional baselines for human motion prediction.","keywords":["human motion prediction","discrete cosine transform","trajectory space","graph convolutional network","learnable graph connectivity","3D pose representation","residual prediction","motion forecasting"],"falsifier":"Run the same network on a motion sequence with sharp, high-frequency joint excursions, such as a fast kick or a stumble, truncate the DCT coefficients as prescribed, and compare MPJPE against the same model with all coefficients retained; if truncation consistently degrades accuracy on such non-smooth motions, the smoothness premise fails.","tokens_in":19258,"feed_emoji":"🏃","tokens_out":6071,"duration_ms":57159,"temperature":0.7,"pith_summary":"Human motion prediction has usually been built on recurrent neural networks, which can accumulate errors and produce jerky transitions. This paper proposes a compact feed-forward alternative: encode each joint's trajectory with Discrete Cosine Transform coefficients, predict residual coefficients for the future, and decode them back to poses. Spatial structure is handled by graph-convolution layers that learn their own connectivity, instead of assuming the skeleton's kinematic tree. The authors report state-of-the-art accuracy on three standard benchmarks, with the best results when training directly on 3D coordinates, and argue that angle-space evaluation is unreliable because different angle sets can describe the same pose.","feed_headline":"DCT trajectories beat recurrent nets at human motion prediction","feed_subtitle":"A compact feed-forward network using smooth trajectory coefficients outperforms RNN baselines on three motion benchmarks.","key_machinery":"The load-bearing object is the trajectory-space DCT representation: each joint's position or angle over time is expanded in Discrete Cosine Transform bases, and the network operates on the resulting coefficients rather than on raw frames. The graph-convolution layer with a learnable weighted adjacency matrix $A^{(p)}$ per layer is the mechanism that lets the network choose which joints exchange information; no pre-defined skeleton graph is used. Residual prediction, obtained by padding the observed sequence with copies of the last frame and predicting only the difference in coefficient space, carries the zero-velocity baseline into frequency space.","core_discovery":"The paper's central claim is that motion prediction can be reformulated as residual prediction in trajectory space: a feed-forward network maps the DCT coefficients of the observed, last-pose-padded trajectory to the coefficients of the true future sequence, and inverse DCT turns the output back into poses. Because human motion in 3D coordinates is smooth, truncating high-frequency DCT coefficients acts as a built-in smoothing prior, preventing the jitter that comes from frame-by-frame regression. The same network treats the body as a fully connected graph with a trainable adjacency matrix per layer, allowing it to discover joint dependencies beyond the kinematic tree. With this design, the paper reports the lowest mean-per-joint position error on all three benchmarks for both short-term and long-term horizons, especially when the model is trained directly on 3D joint positions.","pith_inferences":["The same trajectory-space recipe should transfer to other smooth articulated-motion forecasting tasks, such as hand tracking or animal locomotion, wherever per-joint trajectories are reasonably smooth.","The learned adjacency matrices can be inspected after training to reveal whether the network rediscovers known anatomical couplings or discovers action-specific joint dependencies.","A natural extension is to make the number of retained DCT coefficients adaptive per joint or per action, since the paper's own supplement shows angle trajectories require many more coefficients than 3D trajectories.","The reported angle-space failures suggest a testable hypothesis for future benchmarks: comparing methods in angle space systematically penalizes models whose predictions are geometrically correct but parameterized differently."],"forward_implications":["Recurrent models can be replaced by a feed-forward network with a fixed observation window, removing the need to manage recurrent state and the error accumulation that comes with it.","Training on 3D coordinates gives lower position error than training on joint angles, and angle-space errors can mislead evaluation because different angle sets can yield the same 3D pose; future protocols should report 3D error directly.","Truncating high-frequency DCT coefficients acts as a smoothing regularizer, so the same architecture can trade off detail and jitter by choosing how many coefficients to keep.","A per-layer learned adjacency matrix captures cross-limb dependencies that a fixed skeletal graph would miss, improving accuracy on actions that require synchronized body parts."],"supporting_citations":[{"why":"Supplies the trajectory-space DCT idea the temporal encoding is built on.","marker":"[1]"},{"why":"An earlier feed-forward baseline whose kinematic-tree spatial encoding the paper extends.","marker":"[5]"},{"why":"State-of-the-art adversarial RNN baseline the paper compares against and outperforms on average.","marker":"[9]"},{"why":"Provides the Human3.6M dataset and the MPJPE metric used for 3D evaluation.","marker":"[10]"},{"why":"Introduces the graph-convolution formulation that the paper modifies with learned adjacency.","marker":"[13]"},{"why":"Provides the convolutional sequence-to-sequence baseline and the benchmark protocol for comparison.","marker":"[16]"},{"why":"Establishes the zero-velocity and residual prediction baselines that motivate frequency-space residual learning.","marker":"[17]"},{"why":"Inspires the paper's learned neighborhood structure through attention-based graph methods.","marker":"[22]"}],"fun_headline_variants":["Trajectory-space feed-forward nets beat RNNs in motion prediction","DCT trajectory residual learning tops human motion prediction","Graph-net trajectory model sets SOTA on three motion benchmarks","Feed-forward trajectory residual net beats RNN baselines","Learning joint graphs and DCT smoothness for pose forecasting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result rests on the premise that the temporal trajectory of each joint is smooth enough that a small set of low-frequency DCT coefficients retains the predictive signal, which the paper's own supplement shows holds for 3D coordinates but fails for angle trajectories.","fun_headline_variants_meta":{"raw":{"variants":["Trajectory-space feed-forward nets beat RNNs in motion prediction","DCT trajectory residual learning tops human motion prediction","Graph-net trajectory model sets SOTA on three motion benchmarks","Feed-forward trajectory residual net beats RNN baselines","Learning joint graphs and DCT smoothness for pose forecasting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000624,"raw_usage":{"total_tokens":2904,"prompt_tokens":971,"completion_tokens":1933,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":1852}},"tokens_in":587,"tokens_out":1933,"duration_ms":14708,"temperature":1.0,"reasoning_tokens":1852,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:13:57.265602+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same network on a motion sequence with sharp, high-frequency joint excursions, such as a fast kick or a stumble, truncate the DCT coefficients as prescribed, and compare MPJPE against the same model with all coefficients retained; if truncation consistently degrades accuracy on such non-smooth motions, the smoothness premise fails.","supporting_citations":[{"cited_title":"Nonrigid structure from motion in trajectory space","cited_arxiv_id":null,"evidence_quote":"Supplies the trajectory-space DCT idea the temporal encoding is built on."},{"cited_title":"Black, Danica Kragic, and Hed- vig Kjellstrom","cited_arxiv_id":null,"evidence_quote":"An earlier feed-forward baseline whose kinematic-tree spatial encoding the paper extends."},{"cited_title":"Adversarial geometry-aware human mo- tion prediction","cited_arxiv_id":null,"evidence_quote":"State-of-the-art adversarial RNN baseline the paper compares against and outperforms on average."},{"cited_title":"Human3.6m: Large scale datasets and predic- tive methods for 3d human sensing in natural environments","cited_arxiv_id":null,"evidence_quote":"Provides the Human3.6M dataset and the MPJPE metric used for 3D evaluation."},{"cited_title":"Semi-supervised classifi- cation with graph convolutional networks","cited_arxiv_id":null,"evidence_quote":"Introduces the graph-convolution formulation that the paper modifies with learned adjacency."},{"cited_title":"Con- volutional sequence to sequence model for human dynamics","cited_arxiv_id":null,"evidence_quote":"Provides the convolutional sequence-to-sequence baseline and the benchmark protocol for comparison."},{"cited_title":"Black, and Javier Romero","cited_arxiv_id":null,"evidence_quote":"Establishes the zero-velocity and residual prediction baselines that motivate frequency-space residual learning."},{"cited_title":"Graph at- tention networks","cited_arxiv_id":null,"evidence_quote":"Inspires the paper's learned neighborhood structure through attention-based graph methods."}],"review_version":1}