{"id":"335d61cc-af77-452c-b8d6-53c020b7fbe6","arxiv_id":"2607.15699","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An RL agent that adaptively decides when to accumulate events and when to run tracking inference improves event-based feature tracking on a new dynamic benchmark, but the gains are less consistent on an existing benchmark.","lead":"This paper trains a reinforcement-learning agent to decide when an event camera has collected enough events to run feature tracking, instead of using a fixed time window. On a new dynamic benchmark it reports large gains, but its 'consistently outperforms' claim is contradicted by one existing benchmark and several settings are tuned on the new test set.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reward hyperparameters (λ=0.2, threshold 0.6) are selected on DEFT, the same benchmark used for the headline gains; cross-domain claim needs held-out tuning.","rationale":"The reader's conditional verdict is appropriate. The strongest claim relies on Table 1 and cross-dataset generalization. The most direct threat is that Eq. (2)'s constants were selected on DEFT; because the reward directly shapes when the agent stops, a λ chosen to maximize FA on DEFT can manufacture part of the margin. Table 4 is explicit evidence: λ=0 gives 0.307, λ=0.1 gives 0.600, λ=0.2 gives 0.642, λ=0.3 gives 0.614 on DEFT. The selected λ is not justified by a validation protocol. The fixed baselines, by contrast, use the original authors' intervals; Table 7 shows those intervals are not tuned on DEFT (0.01s gives 0.378 while 0.05s gives 0.384), so the comparison is not apples-to-apples. If the same tuning effort were applied to the baseline interval, the margin would shrink, though likely not vanish. I do not think this is fraud or that the method is worthless; the idea is plausible and the ablations (state design, privileged critic, count-based comparison) are informative. But the central quantitative claim needs the held-out check. I set verdict_should_be UNCHANGED because the reader already assigned CONDITIONAL; my concern is the same category and does not demand rejection. Agreement: partial, since the reader's weakest_assumption also emphasizes patch-state sufficiency, which I see as less central; the test-set tuning is the part I would prioritize.","tokens_in":19928,"tokens_out":13938,"duration_ms":131316,"concrete_test":"Hold out all of DEFT during development. Use the same PPO pipeline (MultiTrack training, frozen BT/DT) to select λ (and, if needed, the 0.6 threshold and L_j scale) on MultiTrack validation or a small held-out split of EC/EDS. Then freeze the chosen constants and evaluate on DEFT, EC, and EDS. Repeat with at least 3 random seeds for DT+Ours and BT+Ours and report mean±std of FA/EFA. If the DEFT gain over the base tracker falls below the best fixed/count heuristic in Table 7, or if the EC result for BT+Ours reverses, the general cross-domain advantage is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is that the policy's reward function is tuned on the same benchmark that is used to report the headline improvements. In Eq. (2), the reward is r = L_j * max(-1, 0.6 - e_i) - λ. The threshold 0.6, the scale L_j=2·l_j, and the penalty λ are free constants. Section 5.1 fixes them, and Table 4 selects λ=0.2 by running the full agent on the DEFT dataset (FA goes 0.307→0.642 as λ varies). The main results in Table 1 are then measured on that same DEFT set. Thus the claim that adaptive RL aggregation 'consistently outperforms' fixed windows is, in part, a claim about a configuration chosen to maximize the evaluation metric on the evaluation set. This matters because the abstract and conclusion generalize to 'diverse motion patterns' and 'hand-crafted rule-based alternatives'; a fair test needs the reward constants to be chosen without access to the target benchmark. The issue is compounded by the absence of seeds/error bars, but the selection-on-DEFT is the more specific threat.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GoStop, an RL-based controller that adaptively decides when to accumulate event data ('go') or to trigger tracking inference ('stop') for event-based feature tracking. The policy is a lightweight CNN+MLP trained with PPO using a reward based on tracking error, with privileged ground-truth information provided only to the critic. The method is evaluated as a plug-in on Deep-EV-Tracker and BlinkTrack, on the newly introduced DEFT dataset and on the established EC and EDS benchmarks. The authors report large gains on DEFT, positive average gains on EC/EDS, and introduce a new dynamic event-based tracking dataset.","tokens_in":20241,"tokens_out":6540,"duration_ms":54588,"significance":"If the empirical claims are robust, this is a useful contribution: it is one of the first works to formulate event accumulation for feature tracking as a sequential decision problem, it demonstrates substantial gains when plugged into two recent trackers, and it releases code and a new dynamic dataset. The reported policy overhead is small (about 0.4 ms per decision versus 5.3 ms for the tracker), and the supplementary experiments across event representations support generality. However, the strength of the claims currently outruns the evidence in several specific ways detailed below.","major_comments":[{"comment":"The abstract and conclusion state that the method 'consistently outperforms' heuristic/fixed-window approaches. This is contradicted by the EC row of Table 2: BT+Ours yields FA 0.819 and EFA 0.811 versus BT's 0.833 and 0.819, i.e., a clear degradation on a standard benchmark. Please restrict the claim to averages or to DEFT, or analyze when adaptive stopping hurts under near-uniform motion.","section":"§5.3, Table 2; Abstract"},{"comment":"The reward hyperparameters λ=0.2, the threshold 0.6, and L_j=2·l_j are selected by experiments on DEFT (Table 4 varies λ on DEFT), and the headline results in Table 1 are measured on the same DEFT benchmark. Because DEFT is introduced in this paper and no separate validation/test split is described, the DEFT gains are not an independent test of generalization. Please select hyperparameters on a held-out split (or on MultiTrack) and then report test results once; also report the sensitivity of EC/EDS performance to λ.","section":"§5.1, Eq. (2); §5.4, Table 4"},{"comment":"No variance or multiple-seed results are reported for the RL policy. PPO is stochastic and the reward function contains free constants, so a single training run can be misleading. Please report mean ± std over at least three seeds for the final method and for the ablations, and indicate whether the gains are statistically significant.","section":"§5.1; Tables 1–7"},{"comment":"The Table 1 caption states 'All methods use only event data,' but BlinkTrack (ref. [82]) is originally an event+image tracker. If an event-only variant is evaluated, this must be stated explicitly and the baseline numbers for BT must be regenerated with that variant; otherwise the comparison to the official BlinkTrack is unfair and the plug-and-play claim is unclear. Please clarify the exact input modality and configuration used for both baselines and the proposed integration.","section":"§5.2, Table 1 caption"}],"minor_comments":[{"comment":"The notation L_j is used in the reward but not defined until Section 5.1. Define L_j and l_j in the methodology, and specify the units of the normalized error e_i (patch dimension P).","section":"§3.2, Eq. (2)"},{"comment":"Please describe the episode termination condition and whether a stop is forced at the end of the clipped 95-event-frame sequence. Since the reward for 'go' is identically zero, the finite-horizon behavior is important for understanding the training objective.","section":"§3.2, §5.1"},{"comment":"The runtime rows are difficult to parse, especially for the count-based strategies and the 'Ours' column. Label each row with the corresponding slicing strategy and report the exact runtime for the adaptive policy separately from the tracker runtime.","section":"Table 7"},{"comment":"The claim of introducing RL to event-based feature tracking 'for the first time' is strong given prior RL-based tracking works cited later in the paper. Please soften or carefully delimit the novelty claim.","section":"§1, §2"},{"comment":"When comparing different event representations, clarify whether the BlinkTrack baseline is retrained for each representation and whether the same pretrained tracker is used for the proposed method. Otherwise the gains may reflect representation-specific retraining.","section":"Supplementary, Table 2"}],"recommendation":"major_revision","confidential_remarks":"The core idea is plausible and the DEFT gains are large, but the abstract-level claim is contradicted by Table 2, and the reward hyperparameters are selected on the same DEFT benchmark used for the headline results. These issues are fixable with additional experiments, explicit validation/test separation, and rephrased claims. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is genuinely new for event-based feature tracking: train an RL agent to decide when to stop accumulating events and trigger tracker inference. That hasn't been done before for online tracking, and the plug-and-play framing is practical. The gains on DEFT are large and mostly per-scene consistent, the policy is lightweight (0.4 ms per decision), and the paper shows it works across multiple event representations. The DEFT dataset itself is a useful contribution for dynamic-motion evaluation. The central weakness is exactly what the stress-test flagged: the reward constants (lambda, the 0.6 threshold, L_j scaling) are selected on DEFT, and the headline numbers are measured on DEFT. Table 4 is an ablation of lambda on that same benchmark, and the best lambda is then used for Table 1. That is genuine test-set tuning, and it undercuts the claim of a cross-domain adaptive advantage. This is not a manufactured flaw - the paper reports the ablation tables plainly, so the evidence is on the table. The authors could fix it by tuning on a MultiTrack validation split or a held-out set of own sequences and then reporting DEFT as held out. There are also smaller issues. The abstract says 'consistently outperforms,' but Table 2 shows BT+Ours worse than BT on EC (FA 0.819 vs 0.833), so that wording is too strong. No multiple-seed variance is reported, which matters for an RL method with a 12-hour training run. The patch-only state is a real limitation: if the feature leaves the local patch during a long 'go' stretch, the policy has no signal to recover. The authors acknowledge in the supplement that adaptive event slicing exists offline, and their online framing is a fair distinction, though the novelty is incremental. The paper is well put together. The ablations on state design, privileged information, and the comparison against search-based static windows show clear thinking and honest reporting. The citation pattern looks fine. The math and the experimental protocol are transparent enough to reproduce, and the authors promise code and dataset. This deserves a serious referee. My recommendation: send it to peer review, but require the authors to (1) tune reward hyperparameters without access to the test benchmark, (2) report multiple seeds with variance, (3) retract or soften 'consistently outperforms,' and (4) release the code and trained models. With those changes, this could be a solid method-level contribution to event-based tracking.","headline":"A useful, mostly solid method paper with a real circularity problem: reward hyperparameters are tuned on DEFT, the same benchmark used for the headline gains, and the 'consistently outperforms' claim is contradicted by Table 2.","tokens_in":797,"tokens_out":940,"would_cite":true,"duration_ms":29381,"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":"An RL agent that chooses when to run the tracker beats fixed event windows for event-camera feature tracking.","keywords":["event camera","feature tracking","temporal aggregation","event accumulation","reinforcement learning","PPO","adaptive control","dynamic motion"],"falsifier":"Hold out a new dataset with motion even faster than DEFT and fix the reward constants by training-set validation only; if the reported gains shrink to the level of the Kalman-filter baseline, the advantage is test-set tuned. Alternatively, track a point that moves more than one patch size within a single 'go' run and check whether the tracker can re-acquire it after the stop action.","tokens_in":19807,"feed_emoji":"⚡","tokens_out":4734,"duration_ms":44867,"temperature":0.7,"pith_summary":"The paper makes the case that the biggest remaining bottleneck in online event-based feature tracking is not the tracker itself but the fixed temporal window used to accumulate events before each inference. It claims that a lightweight reinforcement-learning agent, trained with PPO, can learn a go/stop policy that decides when enough events have accumulated to run the tracker, and that this policy adapts automatically to fast, slow, and abruptly changing motion. The evidence is a plug-and-play module that, mounted on two existing trackers, raises Feature Age and Expected Feature Age substantially on a newly collected dynamic benchmark, and the agent adds only negligible runtime. If right, hand-tuned accumulation windows become an obsolete manual step in event-camera tracking.","feed_headline":"Adaptive timing beats fixed windows for event-camera tracking","feed_subtitle":"A go/stop RL agent raises tracking accuracy on fast, changing motion and drops into existing trackers.","key_machinery":"The adaptive temporal agent: a small CNN+MLP policy mapping the current event representation inside a patch, concatenated with the accumulation length, to a binary go/stop action. It is trained by PPO with a reward that penalizes tracking error at 'stop' actions, scaled by accumulated event count, plus a penalty lambda to discourage over-inference. The critic is given privileged training-only information (ground-truth motion, occlusion, remaining sequence length) to stabilize value estimation.","core_discovery":"The central claim is that adaptive temporal aggregation, framed as a sequential decision problem, replaces fixed heuristic accumulation rules in event-based feature tracking. The agent observes the accumulated event pattern in a patch plus the current window length, and chooses either to keep accumulating ('go') or to run tracker inference ('stop'); a reward based on the resulting tracking error and a penalty on excessive inference trains the policy with PPO. The paper reports that this module, applied without changing the underlying trackers, improves average Feature Age from 0.086 to 0.465 for one tracker and from 0.378 to 0.642 for the other on the new DEFT dataset, and improves both on c","pith_inferences":["The same go/stop formulation could be applied to other event-stream tasks with a similar accuracy-versus-latency trade-off, such as event-based object detection or odometry, where the 'inference moment' is currently fixed by a frame timer.","The DEFT dataset's abrupt motion profiles could become a standard robustness probe for any event tracker, not just those with adaptive windows.","The success of privileged information in the critic hints that RL control of perception may need access to ground-truth difficulty during training; a testable extension is whether a self-supervised difficulty estimate could replace ground-truth motion/occlusion labels."],"forward_implications":["Fixed heuristic event windows become a removable bottleneck: the same accumulated-event representation can be fed to the tracker at adaptively chosen times.","The RL module is plug-and-play: it can be attached to existing tracking networks without retraining them, with negligible cost (~0.4 ms per decision).","Accuracy/efficiency trade-off is learned rather than hand-set: the penalty term automatically balances more frequent inference (accuracy) against less frequent inference (speed).","Robustness to abrupt motion improves, suggesting that event cameras' asynchronous nature can be exploited for online tracking beyond fixed frame rates."],"fun_headline_variants":["RL agent learns when to accumulate events for robust tracking","Adaptive event accumulation via reinforcement learning improves tracking","Go/stop RL agent times event accumulation to boost tracking","Adaptive temporal aggregation with RL outperforms fixed rules","Event tracker's accumulation timing learned by RL, not hand-crafted"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The policy's only view is the raw event pattern in a small patch around the last estimate plus the current wait time, and its reward constants were tuned on the evaluation dataset; if a fast-moving feature leaves the patch while the agent is accumulating, or if deployment motion statistics differ from training, the agent has no signal to recover.","fun_headline_variants_meta":{"raw":{"variants":["RL agent learns when to accumulate events for robust tracking","Adaptive event accumulation via reinforcement learning improves tracking","Go/stop RL agent times event accumulation to boost tracking","Adaptive temporal aggregation with RL outperforms fixed rules","Event tracker's accumulation timing learned by RL, not hand-crafted"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000776,"raw_usage":{"total_tokens":3279,"prompt_tokens":761,"completion_tokens":2518,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":2439}},"tokens_in":505,"tokens_out":2518,"duration_ms":17757,"temperature":1.0,"reasoning_tokens":2439,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T22:32:48.285302+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out a new dataset with motion even faster than DEFT and fix the reward constants by training-set validation only; if the reported gains shrink to the level of the Kalman-filter baseline, the advantage is test-set tuned. Alternatively, track a point that moves more than one patch size within a single 'go' run and check whether the tracker can re-acquire it after the stop action.","supporting_citations":[],"review_version":1}