{"id":"617df380-a67e-4308-9b19-09008b6c183b","arxiv_id":"2412.02395","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A trajectory prediction model that adds hand-crafted group detection and field-of-view based social perception features improves pedestrian forecasting accuracy on some benchmarks.","lead":"This paper introduces GPCC, a pedestrian trajectory prediction model that separates nearby people into walking companions (group members) and unrelated strangers, then feeds both types of social cues into a Transformer. It reports accuracy gains on the ETH-UCY benchmark, though the grouping threshold is hand-tuned on the dataset.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ETH-UCY gains rest on a hand-picked distance threshold (dm=20) selected on that dataset; the same modules give no significant benefit on SDD/nuScenes, so the 'across datasets' claim is not yet supported.","rationale":"The paper's abstract and conclusion claim significant improvements 'across multiple datasets,' but the quantitative support is concentrated in the ETH-UCY ablation (Table 3). I examined the Group method and the Conception module. The Conception module is a hand-designed aggregation of distances, directions, and speeds in three FOV partitions; its benefit is plausible but not the main risk. The Group method's central hyperparameter, dm, is the load-bearing component: Supplementary D explicitly says dm=20 was designed 'based on what we calculate among the ETH-UCY dataset,' and the headline ablation is on that same dataset. Because the grouping decision is a hard threshold, the entire Group feature is a discontinuous function of dm. The paper's qualitative analysis shows the model both merging an unrelated pedestrian and splitting a family member, indicating the threshold is doing real work in the reported results. If dm had been chosen on a held-out split, the reported gains would be more convincing; as reported, there is a legitimate selection-bias concern. This is amplified by Tables S3/S4: on SDD and nuScenes, where the same dm is applied, the full model's advantage over the no-module baseline is negligible (0.2% and <1% FDE), so the claimed cross-dataset generality is not demonstrated. A second, secondary observation is that the v3 baseline (0.24/0.46) is markedly worse than the published MSN results (0.21/0.34), so part of the gap may reflect an under-tuned backbone rather than the proposed modules alone; however, the threshold issue is more directly tied to the paper's central novelty and is explicitly acknowledged in the supplement. A concrete held-out threshold-selection experiment would settle whether the ETH-UCY gain survives without test-set information. If it does, the mechanism is real; if not, the central claim should be substantially weakened. I therefore maintain the reader's CONDITIONAL verdict: the concern is real and addressable, and no change to the verdict is required.","tokens_in":17942,"tokens_out":6359,"duration_ms":69246,"concrete_test":"Re-run the Table 3 ablation on ETH-UCY with threshold selection held out: for each leave-one-out fold, select dm from {5,10,15,20,25,30} using only a validation split of the four training scenes, train the full model with that dm, and evaluate on the held-out scene. Report full-model vs v3 ADE/FDE for each fold. If the 33–58% improvement disappears or shrinks to the level seen on SDD, the headline gain is an artifact of test-set-tuned dm.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim (Table 3) is that GPCC improves ADE/FDE from 0.24/0.46 (v3, both modules disabled) to 0.18/0.29 (v0) on ETH-UCY. This improvement is the primary evidence for the Group and Conception modules. However, Supplementary D states that the Group threshold dm=20 was chosen by inspecting the ETH-UCY dataset itself, and the same threshold is then used to produce the headline results on that dataset. Since membership in Ni_group is a hard cut on the sum of pairwise distances, the grouping assignment—and hence the Group feature—is entirely determined by this single scalar. If dm is set too small, family groups split; if too large, unrelated walkers are merged; the paper's own Fig. 3 shows both failure modes occur. On the other hand, when dm=20 is applied unchanged to SDD (Table S4), the full model is indistinguishable from the no-module baseline (6.39/10.17 vs 6.40/10.17), and on nuScenes (Table S3) the gain is at most about 1% ADE. Thus the mechanism only produces a large, clean benefit on the very dataset used to tune its main hyperparameter. The abstract's claim of 'significant improvements' across multiple datasets is therefore not supported by the reported evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GPCC, a pedestrian trajectory prediction model that augments a Transformer/MSN backbone with two hand-designed interaction modules: a Group method that classifies neighbors as group members via a hard threshold dm on summed pairwise distances over the observation window (Eq. 1), and a Conception module that encodes averaged visual (distance, direction, velocity) and 'acoustic' (rear-distance) cues in FOV partitions (Eqs. 4-10). The fused self/group/conception features are fed to the backbone (Eq. 11). The authors report competitive or state-of-the-art ADE/FDE on ETH-UCY, SDD, nuScenes, and NBA, and an ablation (Table 3) showing large drops on ETH-UCY when both modules are disabled. Qualitative analyses and 'intervention' visualizations are used to argue for explainability.","tokens_in":18209,"tokens_out":7568,"duration_ms":82149,"significance":"If the ETH-UCY result is reproducible, the paper contributes a simple inductive bias—hard grouping by long-term distance—and an interpretable FOV-based perception feature that can be plugged into other backbones; the public code release and the clean ablation design are strengths. However, the evidence for 'significant improvements across multiple datasets' is currently weak: on SDD the full model is indistinguishable from the no-module baseline (Table S4), and on nuScenes the Group method can slightly hurt (Table S3). The central claim therefore rests on a single dataset whose key threshold was manually chosen from that dataset (Supplementary D), so the breadth of the contribution is not yet established.","major_comments":[{"comment":"The threshold dm=20 is selected by inspecting the ETH-UCY family scene (Supplementary D, Fig. S3) and then used to produce the headline ETH-UCY numbers in Table 3. Because Eq. (1) is a hard binary threshold, the entire Group feature—and therefore much of the v0-v3 gap in Table 3—is controlled by this manually chosen scalar. Since zara1 is one of the leave-one-out test folds, this procedure is at best an optimistic estimate of the method's performance on ETH-UCY and provides no out-of-sample evidence for the grouping mechanism. The paper should either select dm on a validation split (or adapt it per dataset/scene) and report the sensitivity of Table 3 to dm, or clearly state that the reported ETH-UCY result is tuned on that dataset.","section":"Supplementary D, Eq. (1), Table 3"},{"comment":"The abstract's claim of 'significant improvements' across multiple datasets is not supported by the supplementary ablations. On SDD, disabling both modules (v3) gives 6.40/10.17 versus 6.39/10.17 for the full model, i.e., essentially no benefit; on nuScenes, removing only the Group method (v2) gives 1.34/2.95 (best-of-5) and 1.07/2.27 (best-of-10), compared with 1.33/2.94 and 1.08/2.27 for v0, so the Group method can slightly hurt; and on NBA, removing only Conception (v1) yields exactly the same numbers as v0 at both horizons. The large modular benefit appears only on ETH-UCY. The authors should either add per-dataset grouping/adaptation that yields consistent gains, or revise the abstract and framing to claim a dataset-specific improvement.","section":"Tables S3 and S4; Abstract"},{"comment":"All ablations are single runs with no error bars, confidence intervals, or significance tests. Some of the differences used to support the method—e.g., 6.39 vs 6.40 ADE on SDD, or 1.08 vs 1.07 ADE on nuScenes best-of-10—may be within seed-to-seed variation of stochastic trajectory prediction training. Please report multiple seeds (at least three) and, if the word 'significant' is retained, provide paired statistical tests.","section":"Sec. 4.3, Tables 3/S3/S4"},{"comment":"The qualitative analysis in Fig. 3 shows the Group method splitting the family in (a2) and excluding the mother in (e2)/(f2), which are precisely the failure modes of a fixed distance threshold. The text interprets these as 'aligned with human judgment,' but the model has no access to semantic group labels, and these examples illustrate that the grouping cue is noisy. The explainability claim should be tempered or supported by a quantitative grouping-quality metric (e.g., precision/recall against annotated group labels).","section":"Fig. 3 / Sec. 4.5"}],"minor_comments":[{"comment":"The 'Acoustic Information Strategy' does not use sound; it is merely the distance of out-of-FOV agents. Please rename it or state clearly that this is a geometric proxy.","section":"Sec. 3.3"},{"comment":"'Self-Intention' is described as an explicit goal or destination, but the implementation is only an encoder of the observed trajectory. Please align the terminology with the actual architecture.","section":"Secs. 1 and 3.1"},{"comment":"The concatenation is written with '...'; specify the exact ordering of the right, left, and rear features.","section":"Eq. (10)"},{"comment":"The 'contribution ratio' is a post-hoc norm-based attribution; its interpretation as a causal contribution should be justified or softened.","section":"Eq. (16)"},{"comment":"There are several typographical issues: 'emdedding' (Sec. 3.4), 'littl0' (Sec. 4.5), and 'EHT-UCY' (Table S5 caption) should be corrected.","section":"Throughout"},{"comment":"GPCC ties with SocialCircle on ETH-UCY (0.18/0.29); the text 'outperforms' should be made precise by naming the specific baselines it beats.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript currently overclaims cross-dataset generality: the only clean, large modular benefit appears on ETH-UCY, and the main hyperparameter dm is manually tuned on that dataset. A revision that treats ETH-UCY as the primary demonstration, adds threshold sensitivity or per-dataset adaptation, and reports multi-seed variability would substantially strengthen the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: GPCC is a plausible step forward for pedestrian trajectory prediction, but the paper's headline claim—significant improvements across multiple datasets—does not survive contact with its own tables. The real gain is on ETH-UCY, and the main hyperparameter (dm=20) is tuned on that dataset.\n\nWhat is genuinely new: the model splits neighbors into group members and non-group members using a long-term distance kernel, then feeds the group feature and a FOV-based perception feature into a transformer backbone. The ablation on ETH-UCY is clean: disabling both modules raises FDE by 58.6%. The qualitative analysis, including the family-group example and the intervention experiments, shows a genuine effort to demonstrate the mechanism, not just a metric dump.\n\nWhere the paper stumbles: the threshold dm is chosen by inspecting ETH-UCY, as acknowledged in Supplementary D, and then applied unchanged to every other dataset. The stress-test note is correct—this undercuts the claim of dataset-independent gains. On SDD, the full model is indistinguishable from the no-module baseline (6.39/10.17 vs 6.40/10.17); on nuScenes the gain is about one percent. The abstract says 'significant improvements' across multiple datasets, but the evidence only supports ETH-UCY. No error bars are reported either, so even the ETH-UCY numbers could partly be noise.\n\nThe authors are honest about some limitations: the group method relies on coordinate distance and sequence length, and they mention that in the conclusion. The overreach in the abstract and main text is the bigger issue. The threshold should be validated on held-out scenes or learned from data, and the dataset-specific claims should be softened or supported with per-scene analysis.\n\nWho this is for: anyone working on group-aware interaction modeling or social FOV features in trajectory prediction. It deserves a serious referee, but the referee should insist on fixing the threshold selection and qualifying the claims before publication. As it stands, it's a solid paper with a promising core that needs revision to justify the conference-level claims.","headline":"GPCC's real gain is on ETH-UCY only, and the group threshold is tuned on that dataset; the paper overclaims 'significant improvements across multiple datasets.'","tokens_in":18759,"tokens_out":3234,"would_cite":false,"duration_ms":31105,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that separating companions from strangers, by a long-term distance kernel and a human-like field of view, is what makes its GPCC model outperform uniform social pooling on pedestrian trajectory benchmarks.","keywords":["pedestrian trajectory prediction","group modeling","long-term distance kernel","field-of-view perception","social interaction","transformer","ablation study","human motion forecasting"],"falsifier":"Run GPCC on ETH-UCY with $d_m$ swept from very small to very large while keeping the Conception module fixed; if no threshold yields the reported gap over the no-group ablation, then the long-term distance kernel is not doing the work credited to it. Separately, if inserting a synthetic stranger inside the FOV does not shift the predicted trajectory away from that stranger, the Conception module's claimed perceptual mechanism fails the intervention test the paper itself introduces.","tokens_in":17739,"feed_emoji":"🚶","tokens_out":6220,"duration_ms":64876,"temperature":0.7,"pith_summary":"The paper claims that pedestrian trajectory prediction improves when the model stops treating every nearby person the same way and instead sorts them into two categories: companions who walk together with the target person, and everyone else, who are perceived through a human-like field of view with visual and acoustic cues. Its GPCC model uses a long-term distance kernel to decide group membership, then fuses the group representation, the perceived unrelated neighbors, and the person's own path into a Transformer-based predictor. On the ETH-UCY benchmarks the full model reports average 0.18/0.29 ADE/FDE, against 0.24/0.46 when both grouping and perception are switched off, with the largest single drop coming when grouping is removed but perception is kept. A sympathetic reading is that explicitly modeling who walks with whom is what carries the gain, because perception of strangers without group context can misdirect the model.","feed_headline":"Group-aware model cuts pedestrian prediction error","feed_subtitle":"Treating companions and strangers differently, with field-of-view cues, beats uniform social pooling on crowded scenes.","key_machinery":"The load-bearing mechanism is the long-term distance kernel function $K(i,j)$, defined as $1$ when the sum of Euclidean distances between agents $i$ and $j$ over all observed frames is at most $d_m$ and $0$ otherwise; it partitions neighbors into the group set $\\mathcal{N}^{group}$ and the rest. The second piece is the Conception module, which splits the surroundings into left, right, and rear partitions relative to the target's moving direction, averages distance, relative direction, and velocity over visible FOV agents and distance over rear agents, and encodes these averages into a perception feature. The fused feature $f^i = \\tanh(W_{fuse} \\operatorname{Concat}(f_{con}, f_{self}, f_{group}))$ carries all three factors into the Transformer backbone, so the identity of group members versus perceived strangers is the object that determines what the predictor attends to.","core_discovery":"On the paper's own terms, the central discovery is that decomposing social influence into three factors, self-intention, long-term stable grouping relations, and in- or out-of-FOV perception of strangers, yields more accurate and more interpretable trajectory forecasts than the common uniform-pooling treatment of all neighbors. GrouP ConCeption (GPCC) computes group membership through a non-learned long-term distance kernel $K(i,j)$ that flags a neighbor as a group member when the sum of pairwise distances over the whole observation window stays below a manually set threshold $d_m=20$; group members' trajectories are encoded jointly with the target's trajectory, while non-members are summarized by partition-averaged distance, relative direction, and velocity inside a 180-degree FOV plus distance-only cues from the rear. The two streams plus the ego trajectory are fused and fed to a Transformer with a best-of-K multi-style generator, trained with plain $\\ell^2$ loss. The paper reports that ablating the group and perception modules raises ETH-UCY FDE by about 58.6% over the full model, that removing only the Group method hurts more than removing both, and that intervention-style visualizations show the model shifting attention and contribution weights when a manual stranger or companion is inserted.","pith_inferences":["The fixed threshold $d_m=20$, chosen by inspecting ETH-UCY, is the clearest extension point: a per-scene or learned threshold could improve transfer, and the paper's own qualitative example (the mother being excluded when she lags) shows that cumulative distance alone cannot capture 'walks with you' semantics.","Because grouping uses only summed distance, it cannot distinguish a companion walking slowly beside you from a stranger who happens to stay near you; adding heading coherence, relative velocity, or appearance-based cues would be a natural testable extension.","The contribution-ratio and intervention visualizations suggest a counterfactual test: if a synthetic stranger is inserted, predicted paths should bend away from that stranger's motion, and if a synthetic companion is inserted, group feature weight should rise; running this at scale would turn the qualitative explainability claim into a falsifiable behavioral prediction.","The Conception module's regional averaging discards per-agent identity; a learned attention over individual out-of-group agents might preserve the human-like FOV prior while recovering finer interactions the paper does not claim to capture."],"forward_implications":["If the reported ablations hold, simply partitioning neighbors by a distance-based group kernel and FOV-based perception can lower prediction error on crowded pedestrian scenes by tens of percent relative to a uniform-interaction baseline.","On the ETH-UCY ablation, disabling the Group method while keeping Conception gives the largest FDE increase (58.6%), implying that perception of strangers is beneficial only when group members are first separated out.","The method transfers to heterogeneous scenes (SDD, NBA, nuScenes) with smaller or different gains, matching the intuition that group structure matters most where pedestrian groups exist and less for independent vehicles.","The 180-degree FOV choice is justified by human vision physiology and is optimal on ETH-UCY, while a wider FOV helps on the NBA set, so the perception module's geometry should be tuned to the interaction field of the target domain.","At 26 ms per batch of 100 targets on an M2 Max, the extra grouping and perception computations remain fast enough for low-latency applications such as autonomous driving."],"supporting_citations":[{"why":"Supplies the multi-style trajectory generation module and backbone design that GPCC wraps with its group and perception features.","marker":"[38]"},{"why":"Provides the Transformer encoder-decoder used as the prediction backbone after feature fusion.","marker":"[33]"},{"why":"A grouping-aware baseline; GPCC is compared against it and it motivates modeling group relations.","marker":"[2]"},{"why":"GroupNet, the main grouping-related baseline and the source of the NBA data split used in supplementary experiments.","marker":"[41]"},{"why":"The closest angle-based social-interaction baseline whose FOV-style representation GPCC extends; also a comparison target.","marker":"[39]"},{"why":"Cited as physiological grounding for choosing the 180-degree human FOV angle in the Conception module.","marker":"[35]"},{"why":"Supplies the best-of-K evaluation and adversarial trajectory-generation convention used in training and metrics.","marker":"[10]"}],"fun_headline_variants":["Group and view split boost pedestrian forecasts","Who you walk with predicts path better than crowd pooling","Separating companions and strangers improves trajectory prediction","Perception-aware grouping refines pedestrian path prediction","Using group ties and view cone lowers prediction error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on the manual threshold $d_m=20$ in the long-term distance kernel: group membership is whatever falls within that distance sum, chosen once from ETH-UCY and applied unchanged elsewhere, so if that threshold is miscalibrated for a scene the Group method's benefit, and the paper's central improvement, collapses.","fun_headline_variants_meta":{"raw":{"variants":["Group and view split boost pedestrian forecasts","Who you walk with predicts path better than crowd pooling","Separating companions and strangers improves trajectory prediction","Perception-aware grouping refines pedestrian path prediction","Using group ties and view cone lowers prediction error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00015,"raw_usage":{"total_tokens":1222,"prompt_tokens":997,"completion_tokens":225,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":170}},"tokens_in":613,"tokens_out":225,"duration_ms":3142,"temperature":1.0,"reasoning_tokens":170,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:29:56.409576+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run GPCC on ETH-UCY with $d_m$ swept from very small to very large while keeping the Conception module fixed; if no threshold yields the reported gap over the no-group ablation, then the long-term distance kernel is not doing the work credited to it. Separately, if inserting a synthetic stranger inside the FOV does not shift the predicted trajectory away from that stranger, the Conception module's claimed perceptual mechanism fails the intervention test the paper itself introduces.","supporting_citations":[{"cited_title":"Msn: multi-style network for trajectory predic- tion","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-style trajectory generation module and backbone design that GPCC wraps with its group and perception features."},{"cited_title":"Attention is all you need","cited_arxiv_id":null,"evidence_quote":"Provides the Transformer encoder-decoder used as the prediction backbone after feature fusion."},{"cited_title":"Learning pedestrian group representations for multi-modal trajectory prediction","cited_arxiv_id":null,"evidence_quote":"A grouping-aware baseline; GPCC is compared against it and it motivates modeling group relations."},{"cited_title":"Socialcircle: Learning the angle-based social in- teraction representation for pedestrian trajectory prediction","cited_arxiv_id":null,"evidence_quote":"The closest angle-based social-interaction baseline whose FOV-style representation GPCC extends; also a comparison target."},{"cited_title":"Handbuch der physiologischen Optik","cited_arxiv_id":null,"evidence_quote":"Cited as physiological grounding for choosing the 180-degree human FOV angle in the Conception module."},{"cited_title":"Social gan: Socially acceptable tra- jectories with generative adversarial networks","cited_arxiv_id":null,"evidence_quote":"Supplies the best-of-K evaluation and adversarial trajectory-generation convention used in training and metrics."}],"review_version":1}