{"id":"a3aa855e-92f7-4cf4-a1fe-a62e8da6da4e","arxiv_id":"2506.09083","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"BakuFlow is a desktop annotation tool that extends YOLOE auto-labeling to support multiple visual prompts per class, and combines label propagation, a live magnifier, and data augmentation.","lead":"BakuFlow is a new semi-automatic image annotation tool that combines a live magnifier, data augmentation, label propagation, and YOLOE-based auto-labeling with multiple visual prompts per class. It targets the labeling bottleneck for video frames and dynamic detection datasets, but the paper reports no measurements that show it actually reduces labeling effort.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central workload-reduction claim is unsupported: no quantitative evaluation is reported, and Section V concedes that YOLOE-based detection accuracy may degrade as class/prompt counts grow, directly undermining the scalability claim.","rationale":"I agree with the reader's weakest_assumption: the load-bearing premise is that the multi-prompt YOLOE auto-labeling core is accurate enough to meaningfully reduce manual labeling effort, and this premise is untested. In fact, the paper's own limitation statement in Section V is stronger than the reader's summary: it says that as the number of classes detected from a single visual prompt increases, the stability and accuracy of YOLOE-based detection may degrade. This directly conflicts with the abstract's characterization of the tool as 'flexible and scalable' for dynamic, real-world datasets. The paper is not internally inconsistent, because it acknowledges the limitation and frames it as future work, but the central efficiency claim is not supported by any experiment or user study. The public repository and demonstration video are real engineering artifacts and count as some evidence that the tool exists, but they do not substitute for a measurement of workload reduction. My proposed benchmark would settle whether the claimed speedup materializes and whether the multi-prompt extension helps or hurts as class and prompt counts scale. Because the rejection rests on absence of evidence combined with an admitted limitation, my stress-test does not change the reader's verdict.","tokens_in":4690,"tokens_out":2235,"duration_ms":27628,"concrete_test":"Run a controlled annotation experiment using the public BakuFlow repository: take a fixed set of 500 video frames with ground-truth boxes (e.g., from MOT17 or a retail planogram-style sequence), and have annotators label them under four conditions: (a) manual-only LabelImg, (b) BakuFlow with label propagation and magnifier but no auto-labeling, (c) BakuFlow with auto-labeling using 1, 3, and 10 visual prompts per class, and (d) BakuFlow with auto-labeling for 2, 5, and 10 classes while holding prompt count fixed. Measure per-frame annotation time, number of manual corrections, and auto-label precision/recall against ground truth. If annotation time or correction counts do not decrease, or if auto-label accuracy falls as prompts or classes grow, the central workload-reduction claim is falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that BakuFlow substantially reduces labeling workload in practice. For that claim to hold, the auto-labeling engine must produce labels that require little manual correction, and the multi-prompt extension must preserve accuracy as the number of classes and prompts increases. Neither condition is measured. Section IV contains only a feature checkmark table; there are no annotation-time measurements, no counts of manual corrections, no precision/recall numbers, and no comparison on a common dataset. Section V explicitly concedes that as the number of classes to be detected from a single visual prompt increases, the stability and accuracy of YOLOE-based detection may degrade. That is an internal admission that the scalability premise is insecure. Furthermore, the paper does not describe how multiple visual prompts per class are fused into YOLOE's visual prompt embeddings; without this detail or an evaluation of the fusion strategy, one cannot determine whether adding prompts genuinely improves generalization or merely dilutes prompt specificity. The workload-reduction claim therefore rests on an unverified assumption about auto-labeling accuracy, and the paper's own limitation statement weakens rather than strengthens that assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper describes BakuFlow, a semi-automatic annotation tool implemented in Python with PyQt5 and OpenCV. The tool provides four main features: a live adjustable magnifier for pixel-precise manual corrections, an interactive data augmentation module, label propagation across consecutive frames, and an auto-labeling module based on a modified YOLOE that supports multiple visual prompts per class. The stated central claim, repeated in the abstract and conclusion, is that these features substantially reduce labeling workload and improve efficiency for object detection and tracking in practical scenarios.","tokens_in":4841,"tokens_out":3668,"duration_ms":41761,"significance":"The open-source release of the tool and its integration of YOLOE-based visual prompting with flexible per-class prompts are commendable and address a practical bottleneck in data annotation. However, the paper does not provide any quantitative evidence for the workload-reduction claim; the evaluation is a feature comparison table without annotation-time measurements, accuracy metrics, or a user study. If the claimed benefits were verified with experiments, the tool could be a useful contribution to the annotation ecosystem, but the current evidence does not establish that.","major_comments":[{"comment":"The evaluation consists solely of a feature checklist; there are no quantitative measurements such as labeling time, number of manual corrections, precision/recall numbers, or a comparison with LabelImg and CVAT on a common dataset. Since the abstract and conclusion assert that BakuFlow 'substantially reduces labeling workload and improving efficiency,' this central claim is unsupported by the reported evidence and must be verified with an empirical study.","section":"Section IV, Table I"},{"comment":"The paper asserts that assigning multiple visual prompts per class improves auto-labeling accuracy and robustness, yet it does not describe how the multiple visual prompt embeddings are fused into YOLOE, nor does it present experiments demonstrating that detection accuracy is maintained or improved as the number of classes and prompts grows. The workload-reduction benefit of the auto-labeling module depends directly on this accuracy premise, so the omission is load-bearing.","section":"Section III.C"},{"comment":"The future-work section concedes that as the number of classes to be detected from a single visual prompt increases, the stability and accuracy of YOLOE-based detection may degrade. This concession directly qualifies the paper's earlier claim of 'flexible and scalable labeling for dynamic, real-world datasets' and should be addressed with empirical measurements, such as detection accuracy versus the number of classes and prompts, instead of being deferred to future iterations.","section":"Section V"}],"minor_comments":[{"comment":"The phrase 'labeling (or annotation) data' is awkward; consider rewording to 'labeling data (i.e., annotation)' or similar for clarity.","section":"Abstract"},{"comment":"The tool name 'V oTT' contains a formatting artifact with a space; it should read 'VOTT' consistently.","section":"Section I"},{"comment":"The final sentence is grammatically incomplete: 'An AI visual platform will be integrated with BakuFlow and explored be for better labeling' appears to be missing a word and should read 'explored for better labeling.'","section":"Section V"},{"comment":"The screenshots in Figure 1 are small and the captions are minimal; larger, more detailed panels would help readers understand the user interface and the operations described in the text.","section":"Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is closer to a software demonstration than a full research paper, and the lack of any quantitative evaluation is a serious concern for a journal-level submission. The central claim of workload reduction is not backed by measurements, and the one admitted limitation in Section V undermines the scalability assertion. I would be willing to reconsider a revised version that includes an empirical user study or benchmark comparing annotation time and label quality against existing tools, along with a description and evaluation of the multi-prompt fusion strategy. Without that evidence, the paper would not meet the journal's bar for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"BakuFlow is a lightweight open-source annotation tool with a live magnifier, data augmentation, label propagation, and a YOLOE-based auto-labeling module extended to take multiple visual prompts per class. That multi-prompt extension is the one genuinely new piece: the original YOLOE accepts a single prompt per class, and the authors changed that to accept several. It is a reasonable idea, and the paper is honest about its limits — Section V concedes that as the number of classes per prompt grows, YOLOE's accuracy and stability may degrade. The GitHub link means the artifact itself is available, which is real.\n\nWhat the paper does well is keep the system description clear and the feature comparison concrete. The comparison table to LabelImg and CVAT is useful, and the discussion of use cases (sequential video frames, planogram compliance) gives context. The writing is straightforward, and the limitation statement is candid.\n\nThe soft spot is large and central: the abstract and conclusion claim the tool 'substantially reduces labeling workload,' but no measurement supports that. Section IV is a checkbox table, not an evaluation. There are no annotation-time numbers, no counts of manual corrections, no precision/recall, no user study, and no comparison on a public dataset. The workload-reduction claim depends entirely on the auto-labeler producing labels that need little fixing, and that is unverified. The stress-test's fusion point also stands: the paper does not explain how multiple visual prompts are combined into YOLOE's embedding, so it is unclear whether adding prompts improves generalization or simply dilutes prompt specificity. Without that detail, the core technical contribution is hard to assess.\n\nNone of this is a mathematical error; it is an evaluation gap in a systems paper. The tool may well be useful to practitioners who want a portable, simple annotator with a YOLOE backend. But as a research claim, the current submission does not support its central assertion.\n\nFor peer review: I would not send this to a full track as is. The better path is to have the authors add a quantitative evaluation — time per frame, accuracy versus manual labels, and a comparison with CVAT or Roboflow on a standard video dataset — and then resubmit. That is a desk reject with clear encouragement, not a burial.","headline":"Open-source annotation tool with a useful multi-prompt YOLOE idea but no evaluation to back the workload-reduction claim.","tokens_in":5388,"tokens_out":2998,"would_cite":false,"duration_ms":35417,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"BakuFlow claims that combining label propagation, a live magnifier, data augmentation, and a multi-prompt YOLOE auto-labeling engine substantially cuts the manual workload of labeling detection and tracking data, especially across video…","keywords":["image annotation","label propagation","auto-labeling","visual prompting","YOLOE","object detection","data augmentation","video frame labeling"],"falsifier":"Label a fixed video twice with BakuFlow, once using a single visual prompt per class and once using multiple prompts per class, with the same set of classes, and compare the auto-generated boxes against a manually corrected ground truth: if multi-prompt precision and recall are not at least equal to single-prompt results, or if per-class accuracy drops as classes are added, the claimed workload reduction fails because the human would have to redraw too many boxes.","tokens_in":4463,"feed_emoji":"🏷️","tokens_out":9146,"duration_ms":83507,"temperature":0.7,"pith_summary":"This paper presents BakuFlow, a semi-automatic labeling tool aimed at object detection and tracking datasets. Its central claim is that combining frame-to-frame label propagation, a live magnifier, interactive data augmentation, and an auto-labeling engine built on the YOLOE visual-prompt detector reduces the manual labeling workload compared with drawing every bounding box by hand. The key extension over the original YOLOE is that annotators can add new object classes and attach any number of visual prompts to each class during a labeling session. That flexibility is meant to capture intra-class variation and keep auto-labels usable on real-world datasets whose contents evolve over time. If the claim is right, teams labeling video frames or large image sets would spend their time verifying and correcting machine suggestions rather than drawing boxes from scratch.","feed_headline":"Labeling tool auto-annotates video frames with multi-prompt YOLOE","feed_subtitle":"Combines frame-to-frame label copying, a pixel-precise magnifier, and per-class visual prompts to cut bounding-box workload.","key_machinery":"The load-bearing mechanism is the modified YOLOE auto-labeling engine. YOLOE is a real-time object detector that uses visual prompt embeddings (VPEs), compact representations of what a class looks like derived from example images, rather than a fixed set of class names. BakuFlow's modification lets the user assign many prompts per class and register new classes mid-session, so the detector's notion of each class can grow with the data. Supporting machinery includes label propagation, which copies selected or all bounding boxes to the next frame; a live magnifier that follows the cursor for fine-grained box adjustment; and interactive augmentation with rotation, brightness, contrast, saturation, and flipping controls to create enriched samples.","core_discovery":"The discovery this paper advances is a design claim: a labeling tool can largely automate bounding-box generation by letting the annotator define each class through multiple example crops, not through a fixed class list. BakuFlow turns user-drawn boxes into visual prompt embeddings, so a few boxes on one frame become prompts that detect the same objects in later frames; new classes can be introduced at any time; and each class can hold several prompts to cover varying appearances. Around that core, the paper contributes three workflow accelerators: label propagation for copying boxes across consecutive frames, a cursor-following adjustable magnifier for pixel-precise corrections, and an augmentation UI for generating enriched training data. The paper's claimed outcome is that this combination substantially reduces labeling workload and improves efficiency for both single-image and cross-image annotation in practical computer vision and industrial scenarios.","pith_inferences":["A controlled trial that measures annotation time per frame for BakuFlow versus fully manual tools would settle the workload-reduction claim, since the paper itself does not report such measurements.","The same multi-prompt idea could transfer beyond detection: the first frame of a tracked object could serve as its visual prompt in later frames, turning the auto-labeler into a tracking assistant.","A testable extension is to measure auto-label precision and recall as the number of classes and prompts per class grows; the paper explicitly anticipates accuracy loss with more classes per prompt, so such a curve would show where manual correction becomes necessary.","The practical payoff depends on how few boxes a human must redraw, not on full automation; even partial auto-labeling can cut cost if the verification step is quick."],"forward_implications":["For video sequences with small inter-frame movement, label propagation reduces repeated box drawing to a copy-and-adjust step.","Because classes are defined by example prompts rather than a fixed classifier, new object types can be added during annotation without retraining the detector.","Multiple prompts per class cover more intra-class variation, which should make auto-proposals more reliable for cluttered or visually diverse scenes.","Exports to YOLO, VOC, and COCO formats keep the resulting labels usable in standard detection and tracking training pipelines."],"supporting_citations":[{"why":"Supplies the YOLOE visual-prompt detector that BakuFlow extends for multi-prompt auto-labeling.","marker":"[3]"},{"why":"Describes an earlier interactive auto-labeling system that frames the AI-assisted workflow this tool builds on.","marker":"[4]"},{"why":"Surveys auto-labeling techniques and establishes the workload-reduction problem the paper addresses.","marker":"[5]"},{"why":"Establishes deep-learning object detection as the application domain for labeled data.","marker":"[1]"},{"why":"Reviews object detection history and motivates the need for efficient annotation at scale.","marker":"[2]"},{"why":"Defines the YOLO detection paradigm underlying the YOLO-family models that YOLOE and BakuFlow target.","marker":"[10]"}],"fun_headline_variants":["Auto-label video frames with per-class visual prompts","BakuFlow: semi-auto labeling with multi-prompt YOLOE","Frame-to-frame label copying accelerates video annotation","Define classes with crops, auto-detect across frames","Multi-prompt YOLOE: add classes and prompts during labeling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed workload reduction rests on the premise, which the paper does not demonstrate experimentally, that YOLOE extended to accept any number of visual prompts per class keeps detection accuracy and stability high as classes and prompts accumulate; the paper itself concedes that accuracy and stability may degrade as more classes are detected from a single prompt.","fun_headline_variants_meta":{"raw":{"variants":["Auto-label video frames with per-class visual prompts","BakuFlow: semi-auto labeling with multi-prompt YOLOE","Frame-to-frame label copying accelerates video annotation","Define classes with crops, auto-detect across frames","Multi-prompt YOLOE: add classes and prompts during labeling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001355,"raw_usage":{"total_tokens":5480,"prompt_tokens":905,"completion_tokens":4575,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":4495}},"tokens_in":521,"tokens_out":4575,"duration_ms":37910,"temperature":1.0,"reasoning_tokens":4495,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:07:52.684954+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Label a fixed video twice with BakuFlow, once using a single visual prompt per class and once using multiple prompts per class, with the same set of classes, and compare the auto-generated boxes against a manually corrected ground truth: if multi-prompt precision and recall are not at least equal to single-prompt results, or if per-class accuracy drops as classes are added, the claimed workload reduction fails because the human would have to redraw too many boxes.","supporting_citations":[{"cited_title":"AI assisted data labeling with interactive auto label,","cited_arxiv_id":null,"evidence_quote":"Describes an earlier interactive auto-labeling system that frames the AI-assisted workflow this tool builds on."},{"cited_title":"Object detection with deep learning: A review,","cited_arxiv_id":null,"evidence_quote":"Establishes deep-learning object detection as the application domain for labeled data."}],"review_version":1}