{"id":"899eaab8-445f-42f7-b544-bd4feb08c057","arxiv_id":"2507.20163","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A visual player-identification module feeding names into an LLM improves identity-aware basketball video captioning; a new 9,726-clip dataset is released.","lead":"This paper presents LLM-IAVC, a system that identifies basketball players from video frames and feeds their names and visual features into a language model to generate captions that mention who did what. It also releases NBA-Identity, a new 9,726-clip annotated dataset, and reports higher captioning scores on two benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The VC-NBA-2022 columns of Table 3 presuppose a player-centric training set that VC-NBA-2022 does not provide; until the exact pipeline is disclosed, the cross-dataset SOTA claim is not reproducible.","rationale":"The reader's conditional verdict matches my analysis. The dataset construction with bounding boxes for NBA-Identity is independent support, and the PIN's 91.4% MCA shows the visual identification subtask is feasible. However, the central generalization claim depends on the second dataset, where the core data prerequisite for the described training flow (bounding boxes for player-centric clips) is absent. This is an internal tension between Table 1, Sec. 3.5, and Supp. E, not merely a disagreement with consensus. It can be resolved by a disclosure and reproducibility check. I therefore keep the verdict unchanged at CONDITIONAL.","tokens_in":19580,"tokens_out":7246,"duration_ms":88160,"concrete_test":"Examine the released code and data preparation scripts for the VC-NBA-2022 experiments. If the scripts construct the VC-NBA-2022 player-centric clip set and PIN labels by any means other than directly using bounding-box annotations that VC-NBA-2022 publicly provides, then the paper must document that modified pipeline and the Table 3 VC-NBA-2022 results must be rerun under the documented pipeline. Specific check: delete every code path that uses a bbox or tracklet file for VC-NBA-2022, then attempt to train the model with only the exact procedure in Sec. 3.5; if the reported CIDEr 150.7 / BLEU-4 38.6 cannot be regenerated, the cross-dataset result is unsupported as reported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that visual-only player identification is sufficient for state-of-the-art identity-aware captioning. That claim is carried by Table 3, including the VC-NBA-2022 benchmark. But the described training flow (Sec. 3.5) requires a player-centric video clip set built from player bounding boxes, and the supplementary (Sec. E) states that the player identification network must be pre-trained with manually annotated player bounding boxes for each sport. Table 1 lists VC-NBA-2022 as having no bounding boxes. The paper never states how the player-centric clip set or PIN identity labels are obtained for VC-NBA-2022, nor whether a separate 286-player PIN is trained. If pseudo-tracklets from SportsMOT were labeled by matching caption names, that is an undocumented and noise-injecting modification of the pipeline; if instead the NBA-Identity-trained PIN was applied, many VC-NBA-2022 players would be out-of-vocabulary because the two label sets (321 vs. 286 players) are not identical. Either way, the reported VC-NBA-2022 numbers cannot be reproduced from the method as written. Removing those numbers leaves only the self-built NBA-Identity benchmark for the SOTA assertion, which weakens the generalizability claim. The core idea is plausible and the dataset is valuable; the issue is a missing, load-bearing experimental detail, not a conceptual flaw.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LLM-IAVC, a basketball video captioning model that extracts player identities from visual player sequences using a player identification network (PIN), combines these identity features with video context through a bidirectional semantic interaction module (BSIM) and a visual context learning module (VCLM), and feeds the resulting multimodal prompt into an LLM decoder. The authors introduce NBA-Identity, a new identity-aware basketball captioning dataset with 9,726 clips, 321 identity-labeled players, player bounding boxes, and 9 major event types, and report quantitative results on NBA-Identity and VC-NBA-2022. The core claim is that visual-only player identification is sufficient to generate identity-aware captions and achieves state-of-the-art performance without external knowledge assistance.","tokens_in":19906,"tokens_out":3754,"duration_ms":45938,"significance":"If the reported results are reproducible, the paper makes two useful contributions: NBA-Identity is a substantial new benchmark for identity-aware sports captioning, and the idea of deriving player identities from visual player-centric clips rather than from external candidate lists is a sensible and potentially generalizable direction. The paper also promises public code and data, which would aid follow-up work. However, the evidence for the central SOTA claim is weakened by an under-specified cross-dataset pipeline, a confounded baseline comparison, and the absence of uncertainty estimates. The dataset itself is the strongest asset; the method-level claim needs firmer experimental support.","major_comments":[{"comment":"The VC-NBA-2022 columns in Table 3 are not reproducible from the method as written. The training flow in Section 3.5 builds the player-centric clip set by querying player sequences derived from bounding boxes, and Table 1 lists VC-NBA-2022 as having no bounding boxes. Supplementary Section E further states that manually annotated player bounding boxes are required to pre-train the player identification network for each sport. The paper never states how player sequences and PIN identity labels were obtained for VC-NBA-2022, such as whether SportsMOT pseudo-tracklets were used, whether a separate 286-player PIN was trained, or whether the NBA-Identity-trained PIN was applied despite the different identity sets. This is a load-bearing missing detail for the cross-dataset SOTA claim.","section":"Table 1, Section 3.5, Supplementary Section E"},{"comment":"The state-of-the-art comparison is confounded by decoder size. The LLM-IAVC variants use GPT-2, Qwen2.5, and Llama3.2 decoders of up to 3B parameters, whereas the listed baselines use much smaller caption decoders. The rows labeled \"without prompt\" show that even without identity information the larger LLM decoders already outperform several baselines on BLEU-4, so part of the improvement may come from decoder capacity rather than the proposed modules. A matched decoder baseline or an ablation that uses the same LLM decoder without the identity prompt is needed to attribute the gains to the player-centric prompt.","section":"Section 4.3, Table 3"},{"comment":"No error bars, variance, or significance tests are reported anywhere in the experimental sections. With test sets of 1,059 clips for NBA-Identity and 786 clips for VC-NBA-2022, differences of one or two CIDEr points between the best and second-best systems may be within run-to-run noise. The authors should report multiple seeds, bootstrap confidence intervals, or a paired significance test, especially for the headline SOTA comparisons in Table 3.","section":"Section 4.1, Section 4.3"},{"comment":"The game-based train/test split does not prevent the same players from appearing in both training and test games, since the dataset covers only 321 players across 40 games. A player appearing in a training-game clip is also likely to appear in a test-game clip, so the PIN may be evaluated on identities already seen during training. This weakens the claim that identities are recognized from a visual perspective rather than memorized. The authors should either construct a disjoint-identity split or report accuracy stratified by whether test players appeared in training games.","section":"Section 4.1"}],"minor_comments":[{"comment":"The text says \"demotes\" where it should say \"denotes\", and Figure 4 contains the typo \"Droupout\" for \"Dropout\".","section":"Section 3.2, Eq. (7)"},{"comment":"The inference flow in Figure 2 says top-k players are selected, and Section 4.2 specifies top-2 for NBA-Identity, but no k value is stated for VC-NBA-2022; specify k for each dataset.","section":"Figure 2 and Section 4.2"},{"comment":"The abbreviations G, Q, and L for GPT-2, Qwen2.5, and Llama3.2 appear in the table but are defined only in the paragraph after the table; define them directly in the table caption.","section":"Table 3"},{"comment":"The sentence \"training and testing per epoch take 221.2s and 115.3s, respectively, with 0.3s per caption generation\" is ambiguous: it is unclear whether the 115.3s testing time includes feature extraction, tracking, and decoding or only decoding, and whether the 0.3s figure covers the full pipeline.","section":"Section 4.2"},{"comment":"The contribution bullet \"LLM-IA VC\" is missing the hyphen between IA and VC; the notation should be unified as LLM-IAVC throughout the paper.","section":"Contributions"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The NBA-Identity dataset is the real contribution here, and the core idea—recognizing player identities from visual features rather than pulling them from a candidate list—is sound. The paper is worth reading for those two things. The LLM-IAVC pipeline (PIN + BSIM + VCLM prompting an LLM) is a reasonable engineering combination, and the ablations show the player identification network carries most of the performance (CIDEr jumps from 66.5 to 95.1 without it). Shipping code and data is also a point in its favor.\n\nThe soft spots are real but addressable. The biggest one is the VC-NBA-2022 evaluation. Table 1 states that dataset has no bounding boxes, but the training flow in Sec. 3.5 requires a player-centric clip set built from bounding boxes, and the supplementary says the PIN must be pre-trained with manually annotated boxes per sport. The paper never says how player sequences and identity labels were obtained for VC-NBA-2022, or whether a separate 286-player PIN was trained. As written, the VC-NBA-2022 columns of Table 3 are not reproducible. That matters because the 'state-of-the-art on VC-NBA-2022' claim is part of the abstract. If those numbers are removed, the SOTA claim rests entirely on the self-built NBA-Identity benchmark.\n\nAlso missing: error bars or significance tests, which is more important because the METEOR and Rouge-L gaps over EIKA are small. The split by game still lets the same players appear in train and test, so identity recognition may be easier than a fully disjoint split would be. And the comparison is partly confounded by decoder size—their best model uses Llama3.2-3B, while KEANet and EIKA use smaller decoders. The GPT-2 variant still beats them on NBA-Identity CIDEr, so the method has legs, but the headline numbers need a controlled comparison.\n\nOne smaller wording point: they claim 'without external knowledge assistance,' but they use a pre-trained LLM, TimeSformer, and SportsMOT. The phrase really means 'no candidate player lists,' which is true, but the current phrasing invites an unfair reading.\n\nBottom line: this deserves a serious referee. The dataset is valuable, the idea is clear, and the listed problems are fixable with disclosure and a few controlled experiments. I'd ask the authors to spell out the VC-NBA-2022 pipeline or drop those columns, and to report variance and a decoder-matched baseline. If they do, I'd be happy to see it in a good venue.","headline":"A useful new dataset and a plausible visual-identification pipeline, but the VC-NBA-2022 results are not reproducible as written.","tokens_in":20426,"tokens_out":2793,"would_cite":true,"duration_ms":31349,"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":"Player names can be read from broadcast basketball video itself, and that is enough to generate identity-aware captions without external game notes.","keywords":["video captioning","identity-aware captioning","basketball videos","player identification","large language models","multimodal prompt","NBA-Identity dataset","player-centric clip set"],"falsifier":"Run the released code on VC-NBA-2022 using only SportsMOT tracks, without any ground-truth bounding boxes: if the identity-aware CIDEr falls from the reported 150.7 toward the video-only level of about 103.1, then the state-of-the-art result depends on annotation access not supplied by that dataset.","tokens_in":19400,"feed_emoji":"🏀","tokens_out":6116,"duration_ms":67338,"temperature":0.7,"pith_summary":"This paper tries to establish that the missing piece in identity-aware sports video captioning is visual player identification rather than extra game knowledge. It proposes LLM-IAVC, which recognizes player names from the video frames themselves, bundles those names with video context into a multimodal prompt, and lets a large language model produce captions such as \"D. Murray makes 2-pt jump shot (assist by W. Carter)\". To train and test this, the authors build NBA-Identity, a dataset of 9,726 basketball clips with player bounding boxes and 321 player identities. They report that the visual-identification prompt lifts CIDEr from 66.5 to 105.3 for the Llama3.2-3B variant on their benchmark and reaches 150.7 CIDEr on the existing VC-NBA-2022 dataset, surpassing methods that rely on external candidate lists. If correct, the result means accurate player naming does not require external statistics or rosters; the visual signal is sufficient.","feed_headline":"Player names read from video beat knowledge-based captioning","feed_subtitle":"New 9,726-clip NBA-Identity benchmark plus a visual player-ID prompt lets an LLM name who shot, rebounded, and assisted.","key_machinery":"The load-bearing mechanism is the player-centric multimodal prompt. Training starts from a player-centric video clip set: every caption names the players involved, and the dataset's bounding boxes let the paper crop each named player into a sequence of images; a player identification network (PIN), a TimeSformer backbone plus a linear classification head, learns to map those sequences to one of 321 player names. The bidirectional semantic interaction module (BSIM) then exchanges information between the video feature and the top-2 players' visual features with cross-attention in both directions, and the visual context learning module (VCLM) uses 32 learnable query vectors to condense the video into context tokens. The four token groups, video context, video features, player visual features, and player-name text embeddings, are projected into the LLM's embedding space and concatenated as a prompt, so the decoder sees both who is on screen and what is happening. At inference, SportsMOT supplies player tracks, the PIN names and ranks them, and the top-2 are fed forward the same way.","core_discovery":"The paper's central claim is that player identities are visually discoverable in broadcast basketball clips, and a network that learns them from a player-centric clip set can supply an LLM with everything needed to generate identity-aware captions without external knowledge assistance. The discovery is that organizing training clips around individual players, using captions and bounding boxes to link each player to his clips, makes it possible to train a player identification network that achieves 91.40% classification accuracy among 321 players, and that injecting the top-2 identified players' visual features and name embeddings, together with learned video context, into the LLM prompt is what accounts for the large gains in captioning quality. On NBA-Identity the full model reaches 105.3 CIDEr and 18.8 BLEU-4, and on VC-NBA-2022 it reaches 150.7 CIDEr and 38.6 BLEU-4, the best scores in the comparison; the ablation shows that dropping the player identification network reduces CIDEr from 105.3 to 66.5.","pith_inferences":["Left implicit by the paper: the NBA-Identity split by game does not guarantee generalization to never-seen players, since the same 321 players can appear in both training and test games; a held-out-identity split would be a harder and more informative test of visual recognition.","The VC-NBA-2022 comparison is central to the claim, so one easy check is to rerun inference using only SportsMOT tracks and verify whether the reported 150.7 CIDEr persists; a large drop would show the identity prompt depends on annotation access rather than purely visual recognition.","The paper's own admission that shot-distance estimation is weak suggests a natural extension: add a regression head for shooting distance and feed it as an extra prompt token, which would directly address the observed 6-ft versus 4-ft errors without changing the architecture."],"forward_implications":["Identity-aware sports captions can be generated without external game notes or candidate rosters, because player names are read from the video itself.","Feeding player-name tokens plus video-context tokens as a multimodal prompt to an LLM improves naming accuracy and CIDEr by large margins, with the Llama3.2-3B variant gaining 38.8 CIDEr points on NBA-Identity when the prompt is added.","A benchmark with per-player bounding boxes, NBA-Identity, is sufficient to train a visual player identification network, so datasets without identity annotations can be augmented by a similar player-centric pipeline.","The same player-centric prompt design should transfer to other sports after retraining the player identification network on that sport's manually annotated bounding boxes.","Because the player identification network is the main source of name accuracy, improving jerse recognition through player numbers and jersey colors is a direct path to further captioning gains."],"supporting_citations":[{"why":"Supplies the TimeSformer visual backbone used both for video encoding and for the player identification network.","marker":"[3]"},{"why":"Supplies the SportsMOT multi-object tracker that extracts player sequences at inference.","marker":"[8]"},{"why":"Defines the VC-NBA-2022 identity-aware basketball captioning benchmark and the KEANet candidate-list baseline that the paper must beat.","marker":"[50]"},{"why":"Provides the EIKA extra-information baseline, the strongest comparison method using a candidate player list.","marker":"[51]"},{"why":"Llama3.2 is the decoder variant achieving the best reported performance, showing the prompt works with a modern lightweight LLM.","marker":"[24]"},{"why":"Qwen2.5 variants demonstrate that the multimodal prompt gains hold across different LLM decoders.","marker":"[42]"}],"fun_headline_variants":["Visual player IDs boost basketball captioning accuracy","Player names from video, not external data, improve captions","Player-centric prompts lift LLM caption quality on NBA clips","9,726 clips teach LLM to name players in video captions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported VC-NBA-2022 gains assume the player-centric pipeline can be run on that dataset even though it is listed as having no player bounding boxes; the paper never states how player sequences are obtained there.","fun_headline_variants_meta":{"raw":{"variants":["Visual player IDs boost basketball captioning accuracy","Player names from video, not external data, improve captions","Player-centric prompts lift LLM caption quality on NBA clips","9,726 clips teach LLM to name players in video captions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000261,"raw_usage":{"total_tokens":1634,"prompt_tokens":1030,"completion_tokens":604,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":646,"completion_tokens_details":{"reasoning_tokens":536}},"tokens_in":646,"tokens_out":604,"duration_ms":7182,"temperature":1.0,"reasoning_tokens":536,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:41:42.136331+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released code on VC-NBA-2022 using only SportsMOT tracks, without any ground-truth bounding boxes: if the identity-aware CIDEr falls from the reported 150.7 toward the video-only level of about 103.1, then the state-of-the-art result depends on annotation access not supplied by that dataset.","supporting_citations":[{"cited_title":"Qwen2.5: A party of foundation models, 2024","cited_arxiv_id":null,"evidence_quote":"Qwen2.5 variants demonstrate that the multimodal prompt gains hold across different LLM decoders."},{"cited_title":"Is space-time attention all you need for video understanding? In ICML, page 4, 2021","cited_arxiv_id":null,"evidence_quote":"Supplies the TimeSformer visual backbone used both for video encoding and for the player identification network."},{"cited_title":"Sportsmot: A large multi-object tracking dataset in multiple sports scenes","cited_arxiv_id":null,"evidence_quote":"Supplies the SportsMOT multi-object tracker that extracts player sequences at inference."},{"cited_title":"A simple yet effective knowledge guided method for entity-aware video captioning on a basketball benchmark","cited_arxiv_id":null,"evidence_quote":"Defines the VC-NBA-2022 identity-aware basketball captioning benchmark and the KEANet candidate-list baseline that the paper must beat."},{"cited_title":"Eika: Explicit & im- plicit knowledge-augmented network for entity-aware sports video captioning","cited_arxiv_id":null,"evidence_quote":"Provides the EIKA extra-information baseline, the strongest comparison method using a candidate player list."},{"cited_title":"Llama 3.2 quantized models, 2024","cited_arxiv_id":null,"evidence_quote":"Llama3.2 is the decoder variant achieving the best reported performance, showing the prompt works with a modern lightweight LLM."}],"review_version":1}