{"id":"1eff7d50-22c7-443f-9acd-c2b48f05aa24","arxiv_id":"1908.00716","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"EnEx2 is a two-camera entry-exit surveillance dataset with a bounding-box-containment rule for labeling tracks as entry, exit, or pass-by.","lead":"This paper introduces EnEx2, a two-camera video dataset of people entering and leaving a camera-forbidden room, along with a simple baseline that classifies each person's movement as entry, exit, or passing by. The value is a benchmark for privacy-respecting surveillance outside toilets, changing rooms, and other private areas where cameras cannot be placed inside.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed benefit of the second camera is not supported by the experiments as reported; Table 2 is a confounded cross-dataset comparison rather than a matched single-versus-two-camera ablation.","rationale":"The reader's weakest assumption identifies the pseudo-annotated ground truth as the central vulnerability; that is a genuine and important concern because it attacks the reliability of the numbers themselves. The stress-test pass, however, locates the more direct weakness in the inference from those numbers: even if the EnEx2 labels were perfectly accurate, Table 2 would not demonstrate the importance of a second camera. EnEx and EnEx2 differ in camera placement, scene geometry, subjects, recording duration, and annotation workflow, and the reported accuracy gaps are only 0.4 to 1.8 percentage points with no significance testing. A controlled ablation on matched data is necessary to support the causal claim in Section 5. The paper also lacks error analysis and any independent validation of the auto-generated tracks, so both the label quality and the comparability of the two datasets are unverified. Because the deficiency is an evidential gap that could be fixed by releasing the data with human-verified labels and running a single-versus-two-camera ablation, the conditional verdict remains appropriate and unchanged.","tokens_in":5644,"tokens_out":5355,"duration_ms":56922,"concrete_test":"Run the same EnEx2 videos through the pipeline in two configurations: (a) entrance-view only, with no use of the opposite camera; and (b) two-camera configuration, where the opposite view is used to repair broken or occluded tracks as described in Section 3.2. Keep the detector, tracker, entrance rectangle, event rule, and scoring procedure identical across the two configurations, and score against human-verified event labels on the same clips. If configuration (b) does not beat configuration (a) by a margin larger than annotation noise on held-out clips, the Section 5 claim should be removed or substantially softened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5 concludes that 'High Entry-Exit event detection accuracy on EnEx2 dataset demonstrates the importance of a second camera for Entry-Exit Surveillance.' For this causal claim to hold, the comparison must isolate camera count: the same scene, subjects, detector, tracker, annotation procedure, and evaluation protocol should be run in single-camera versus two-camera modes. The paper provides no such experiment. Table 2 only compares EnEx (single camera, different location, different annotation process) with EnEx2 (two cameras, different scene, pseudo-annotations), and the reported gaps are small: 96.2 versus 94.4 for Entry, 98.1 versus 96.4 for Exit, and 98.7 versus 98.3 for Just appeared. With no error bars, no significance test, and no matched data, those differences cannot be attributed to the second camera. The pseudo-annotation issue identified by the reader compounds the problem: if the EnEx2 event labels are generated from the same HOG/Kalman tracks that the detector consumes, high event accuracy may reflect the labeling rule agreeing with itself rather than robust perception. Neither the dataset nor the evaluation code is released, so the cross-dataset difference cannot be independently audited. The central claim is therefore unsupported as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces EnEx2, a two-camera video dataset for Entry-Exit Surveillance, and a simple rule-based event detection method that labels tracks as Entry, Exit, or Just appeared based on whether the first and last bounding boxes lie within a manually marked entrance rectangle. Baseline people-detection and event-detection results are reported for EnEx2 and three existing datasets (EnEx, CAVIAR, PAMELA-UANDES). The paper concludes that the high accuracy on EnEx2 demonstrates the importance of using a second camera.","tokens_in":5891,"tokens_out":6556,"duration_ms":57768,"significance":"If the dataset were released and independently validated, EnEx2 could serve as a useful benchmark for privacy-respecting monitoring of private-area doorways, and the paper addresses a real application gap. The paper is transparent about EnEx2 being pseudo-annotated and makes a clear statement of the problem and the proposed event taxonomy. However, the reported results are not sufficient to establish the central claim about camera count, and the lack of dataset release and independent annotation severely limits the contribution's immediate utility.","major_comments":[{"comment":"The conclusion that high accuracy on EnEx2 demonstrates the importance of a second camera is unsupported. Table 2 compares EnEx (single camera, different location, different annotation procedure) with EnEx2 (two cameras, different scene, pseudo-annotations); there is no matched experiment that runs the same scene, subjects, detector, tracker, and annotation protocol in single-camera versus two-camera modes. The reported differences (Entry 96.2 vs 94.4, Exit 98.1 vs 96.4, Just appeared 98.7 vs 98.3) are small, and with no error bars or significance tests they cannot be attributed to camera count.","section":"Section 5, Table 2"},{"comment":"The EnEx2 ground-truth tracks are generated using the same HOG-based people detector [6] (with background subtraction [5] and Kalman filters [8]) that is later used in the evaluation pipeline. This creates a circularity: the reported EnEx2 detection and event-detection scores in Tables 1 and 2 partly measure how well the detector agrees with its own pseudo-labels rather than how well it matches independent ground truth. The paper labels the dataset 'pseudo-annotated' but provides no manual validation or independent test set to break the circularity.","section":"Section 2.1 and Section 4"},{"comment":"The event detection method relies on the assumption that the first and last bounding boxes of each track are correctly detected and that the entrance rectangle E is accurately marked. No analysis is provided for cases of occlusion, track fragmentation, or bounding-box drift near the entrance, which can change the B ⊂ E decision and hence the event label. The two free parameters (E coordinates and the unassigned-track frame threshold) are not specified in the paper, so the reported numbers cannot be reproduced.","section":"Section 3.2, Table 2"},{"comment":"The evaluation methodology is under-specified. Tables 1 and 2 give single performance numbers without error bars, track counts, confusion matrices, or a definition of the reported accuracy (e.g., per-track or per-frame). It is unclear how many video sequences from each dataset were used, how events were defined, and whether the results are averaged over runs. Without this information, the results cannot be assessed statistically or reproduced.","section":"Section 4"},{"comment":"The paper's main deliverable—the EnEx2 dataset—is not released; no URL, download link, or access instructions appear. As a benchmark contribution, the dataset must be available for independent verification and comparison. Additionally, the claim that two cameras improve re-identification accuracy is not evaluated at all; the paper only reports detection and event-detection results, not re-identification results, despite listing re-identification as a motivation.","section":"Section 1.2 and Section 5"}],"minor_comments":[{"comment":"The paper contains several incomplete references to figures: 'as shown in Figure .' and 'the view in the figure 3b' do not correspond to any labeled figure in the text.","section":"Section 2.1"},{"comment":"The notation for bounding box inclusion is inconsistent: sometimes it appears as 'B⊂ E' and sometimes as 'B⊂E'; please standardize the math notation.","section":"Section 3.2"},{"comment":"The term 'deeply learned people detection method' for HOG is inaccurate; HOG features are not deep learning.","section":"Section 4"},{"comment":"The dataset name is written inconsistently: 'CA VIAR' in Table 1 vs 'CAVIAR' in Section 4. Please unify the spelling.","section":"Table 1 and Section 4"},{"comment":"The reference list contains errors: reference [2] lists the author as 'V , Vinay Kumar' instead of 'Vinay Kumar V', and reference [12] has a missing space in 'PAMELA- UANDES'.","section":"References"},{"comment":"The term 'pseudo-annotated' is used without explaining what pseudo-annotation entails; given the circularity concern, this should be clarified in the methodology.","section":"Abstract and Section 5"}],"recommendation":"major_revision","confidential_remarks":"The circularity between the ground-truth generation and the evaluation detector is the most serious issue; if the authors cannot re-annotate EnEx2 independently or provide a matched single-vs-two-camera experiment, the paper should likely be rejected. The dataset release is also essential for the paper as a benchmark contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the paper gives you something real: EnEx2, a two-camera entry-exit surveillance dataset with track-level ground truth, plus baseline numbers for a straightforward containment-based event detector. That is a useful step for a small subfield. The authors also run their method on EnEx, CAVIAR, and PAMELA-UANDES, which gives some comparability. The idea itself is not deep, but the dataset is the contribution.\n\nThe soft spots are significant. The ground truth for EnEx2 is pseudo-annotated using HOG detections plus background subtraction and Kalman filtering (Section 2.1), and evaluation uses the same HOG-based detector (Section 4). So the people detection and event detection scores partly measure the system against its own output. That is a circularity problem the paper never addresses. On top of that, Tables 1 and 2 report single numbers with no error bars, no track counts, no confusion matrices, so we have no sense of variance or where failures happen.\n\nThe bigger issue is the conclusion: 'High Entry-Exit event detection accuracy on EnEx2 dataset demonstrates the importance of a second camera.' That is not supported. EnEx2 is a different scene, different camera setup, different annotation procedure, and it is compared against EnEx which is single-camera. There is no matched experiment with the same scene and subjects in single- versus two-camera mode. The accuracy gaps in Table 2 are a couple of points, well within noise given the lack of error bars. The third column 'Just appeared' is basically the same. So the title claim overreaches.\n\nA smaller point: EnEx2 is not released, so the dataset's utility is currently hypothetical. We cannot verify the pseudo-annotations or rerun the numbers.\n\nAll that said, the paper is not incoherent. The spatial transition rule is simple and the four-state formulation is clear. The authors are honest that it is pseudo-annotated. It deserves a serious referee, but only with major revision: release the data or at least independent annotations, add error analysis, and either run a matched single-vs-two-camera ablation or remove the causal claim about the second camera.\n\nWho is this for? Researchers working on entry-exit surveillance or privacy-respecting monitoring of doorways. For that niche, the dataset is potentially useful if made available. For general CV, not much. I'd send it to peer review with a request for major revision rather than desk-rejecting it flat.","headline":"A useful but thin two-camera dataset with an overclaimed advantage: the paper does not actually show that the second camera helps.","tokens_in":6434,"tokens_out":2124,"would_cite":false,"duration_ms":20924,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper establishes that two outward-facing cameras can detect who enters and exits a camera-forbidden private area, introducing the EnEx2 benchmark with entry-exit accuracy above 96 percent.","keywords":["entry-exit surveillance","private areas","two-camera dataset","event detection","people tracking","person re-identification","EnEx2","camera-forbidden areas"],"falsifier":"Take a held-out set of EnEx2 clips where people overlap at the doorway, have human annotators label each appearance as Entry, Exit, or Just-appeared, and compare their labels with those produced by the bounding-box-subset rule; systematic disagreement on those clips would show that the reported accuracy depends on the auto-generated tracks rather than on the rule itself.","tokens_in":5429,"feed_emoji":"🎥","tokens_out":5909,"duration_ms":53706,"temperature":0.7,"pith_summary":"The paper is trying to establish that security cameras can monitor people entering and exiting private areas such as toilets and changing rooms without recording what happens inside, by placing two cameras outside and tracking each person until they cross the doorway. To make this measurable, it introduces EnEx2, a two-camera dataset of laboratory door crossings, with auto-generated tracks provided as ground truth. It then proposes a spatial transition rule in which an event is classified as Entry, Exit, or Just-appeared depending on whether the first and last bounding boxes of a person's track lie inside a manually marked entrance rectangle. On EnEx2 the rule reaches 96.2 percent Entry, 98.1 percent Exit, and 98.7 percent Just-appeared accuracy, which the paper takes as evidence that a second camera helps entry-exit surveillance.","feed_headline":"Two cameras lift doorway event detection to 96-98 percent","feed_subtitle":"EnEx2 places two cameras outside a private doorway so entry, exit, and passersby are classified without recording inside.","key_machinery":"The central object is the entrance rectangle $E$, a manually marked region in the camera view that stands for the doorway of the private area, together with the subset test $B \\subset E$ applied to a track's first and last bounding boxes. This test converts tracking into event classification. It is supported by a pipeline of background subtraction, HOG-based people detection, and Kalman-filter tracking that generates the tracks, and by the two-camera EnEx2 setup whose spatial coordinates are calibrated so that the opposite view can recover individuals occluded at the doorway.","core_discovery":"The central claim is that entry-exit surveillance can be reduced to a geometric test on track endpoints. For each track, the initial bounding box state is classified as EA (entered the scene from outside) or EX (exited from the private area) by checking whether the box is a subset of the entrance rectangle E; the final state is classified as EN (entered the private area) or EA (exited the scene to outside) in the same way. The ordered pair of states then determines the event: (EA, EA) is Just appeared, (EA, EN) is Entry, (EX, EA) is Exit, and (EX, EN) is Re-entry, also treated as Just appeared. On the new EnEx2 dataset, with two oppositely placed cameras, this rule yields 96.2 percent Entry, 98.1 percent Exit, and 98.7 percent Just-appeared accuracy, above the single-camera EnEx baseline, which the paper attributes to the second view resolving doorway occlusions.","pith_inferences":["The event rule reads only the first and last frames of a track, so a natural extension is to classify events from the whole trajectory relative to the entrance rectangle, which could survive momentary detector failures at the doorway.","Because the two camera views are spatially calibrated, EnEx2 could also serve as a testbed for multi-view person re-identification and track association, not just event detection.","For real deployments the system stores no imagery from inside the private area, so the remaining privacy exposure is the exterior cameras themselves; a useful extension would add a privacy metric for bystanders caught by the outside view."],"forward_implications":["If the two-camera advantage holds, future entry-exit surveillance systems can rely on exterior cameras only, preserving privacy inside the private area while still logging who enters and exits.","EnEx2 provides a benchmark that researchers can use to compare entry-exit event detection and person re-identification after a privacy interval.","The spatial-transition rule is dataset-agnostic: the paper reports that the same rule also works on CAVIAR corridor sequences and PAMELA-UANDES metro boarding sequences by treating shop doors and train doors as entrance rectangles.","Entry-exit event detection is an extension of tracking, so improvements in people detection and tracking at doorways should directly improve event accuracy."],"supporting_citations":[{"why":"It defines the original single-camera EnEx entry-exit surveillance benchmark that EnEx2 extends and that the proposed method is compared against.","marker":"[2]"},{"why":"It provides the HOG-based people detector used to produce the tracks in the EnEx2 ground truth and in the event-detection pipeline.","marker":"[6]"},{"why":"It supplies the adaptive Gaussian mixture background subtraction used to separate moving people before detection.","marker":"[5]"},{"why":"It provides the Kalman-filter tracking used to follow each detected individual across frames and form tracks.","marker":"[8]"},{"why":"It supplies the CAVIAR corridor sequences used as an additional evaluation set for the proposed event-detection rule.","marker":"[11]"},{"why":"It supplies the PAMELA-UANDES metro sequences used as a second additional evaluation set.","marker":"[12]"}],"fun_headline_variants":["EnEx2 dual-cam test yields 96% entry, 98% exit accuracy","Second camera outside the door clears occlusions for 98% exit","Doorway events from track endpoints: 96-98% accurate, no inside camera"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that whenever someone crosses the doorway, the people detector and tracker produce a complete, correctly placed bounding box in exactly the frame of the crossing, because the event label is simply whether that box falls inside the manually marked entrance rectangle.","fun_headline_variants_meta":{"raw":{"variants":["EnEx2 dual-cam test yields 96% entry, 98% exit accuracy","Second camera outside the door clears occlusions for 98% exit","Doorway events from track endpoints: 96-98% accurate, no inside camera"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000572,"raw_usage":{"total_tokens":2703,"prompt_tokens":945,"completion_tokens":1758,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":1690}},"tokens_in":561,"tokens_out":1758,"duration_ms":14185,"temperature":1.0,"reasoning_tokens":1690,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:34:46.115659+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out set of EnEx2 clips where people overlap at the doorway, have human annotators label each appearance as Entry, Exit, or Just-appeared, and compare their labels with those produced by the bounding-box-subset rule; systematic disagreement on those clips would show that the reported accuracy depends on the auto-generated tracks rather than on the rule itself.","supporting_citations":[{"cited_title":"In: Chaudhuri, B.B., Nakagawa, M., Khanna, P., Kumar, S","cited_arxiv_id":null,"evidence_quote":"It defines the original single-camera EnEx entry-exit surveillance benchmark that EnEx2 extends and that the proposed method is compared against."},{"cited_title":"Fast Feature Pyramids for Object Detection","cited_arxiv_id":null,"evidence_quote":"It provides the HOG-based people detector used to produce the tracks in the EnEx2 ground truth and in the event-detection pipeline."},{"cited_title":"Region-of-Interest Tracking Based on Keypoint Trajectories on a Group of Pictures,","cited_arxiv_id":null,"evidence_quote":"It provides the Kalman-filter tracking used to follow each detected individual across frames and form tracks."},{"cited_title":"5 (6 pp.)-5 (6 pp.), DOI: 10.1049/cp.2018.1281","cited_arxiv_id":null,"evidence_quote":"It supplies the CAVIAR corridor sequences used as an additional evaluation set for the proposed event-detection rule."},{"cited_title":"Video Object Segmentation and Tracking: A Survey","cited_arxiv_id":"1904.09172","evidence_quote":"It supplies the PAMELA-UANDES metro sequences used as a second additional evaluation set."}],"review_version":1}