{"id":"623148ae-84f8-4cfd-ae92-2e77eb0c4e07","arxiv_id":"2607.08402","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A five-stage pipeline using Roop face-swapping anonymizes pedestrians in Egyptian street images while preserving gaze and expression cues for AV intention models.","lead":"The paper builds a five-stage face-swapping pipeline that hides pedestrian identities in street images while keeping facial cues used by autonomous-vehicle models. It is aimed at privacy-safe release of urban driving datasets such as Egypt’s Egy-DRiVeS without destroying training utility.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"Usability for full AV intention/trajectory models rests only on four proxy metrics plus one looking/not-looking extractor, without end-to-end evaluation.","rationale":"The reader’s weakest_assumption correctly isolates the softest link in the strongest claim: proxy metrics and a single binary feature check do not guarantee that full intention/trajectory models retain utility after anonymization. The remainder of the paper is a competent engineering assembly of existing detectors, restorers and swappers; qualitative robustness of Roop on veils and occlusions is informative, and no internal contradiction or calculation error appears. Because the load-bearing gap is already flagged and the recommended remedy (stronger downstream evaluation) matches the reader’s condition for acceptance, the CONDITIONAL verdict and MODERATE confidence stand without adjustment.","tokens_in":8808,"tokens_out":540,"duration_ms":17072,"concrete_test":"Select a labeled subset of JAAD or Egy-DRiVeS frames that contain looking/crossing annotations; run the full five-stage pipeline to produce anonymized counterparts; retrain a standard intention predictor (e.g., the modular ROS architecture from the authors’ prior work or a simple CNN+LSTM baseline) on original vs. anonymized data under identical splits and hyperparameters; if the drop in accuracy or F1 exceeds 5 %, the preservation claim for full-model utility does not hold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the five-stage Roop pipeline preserves facial attributes (expressions, pose, gaze) sufficiently for usability in pedestrian-intention and trajectory-prediction models on Egy-DRiVeS imagery is supported solely by (i) mean landmark difference, blendshape difference, identity cosine similarity and gaze-vector cosine similarity computed on a small set of high-quality close-ups (Table II) and (ii) a qualitative check that a pretrained looking/not-looking extractor yields the same binary labels before/after anonymization on a handful of clearly visible pedestrians (Section III-E / Fig. 5). No intention or trajectory model is ever retrained or evaluated on the anonymized frames, so the leap from these proxies to retained training utility remains untested. The paper itself notes that facial cues are the main features required by such models (Introduction), yet never measures whether the residual differences after swapping (e.g., Roop’s higher identity cosine of 0.20 vs Ghost-v2’s 0.14) or the CodeFormer resize-and-restore step degrade downstream accuracy.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript proposes a five-stage pipeline (YOLOv11 pedestrian detection, SCRFD face detection, CodeFormer restoration, face swapping, OpenCV Poisson blending) that anonymizes pedestrians in ITS/AV street imagery via face swapping while aiming to retain facial attributes (expression, head pose, eye gaze) required for intention and trajectory models. Roop and Ghost-v2 are compared; Roop is selected for better structure/expression preservation and robustness on occlusions and veiled faces typical of the Egy-DRiVeS dataset. Evaluation uses landmark difference, blendshape difference, identity cosine similarity and gaze-vector cosine similarity (Table II) plus a qualitative looking/not-looking check (Fig. 5).","tokens_in":9095,"tokens_out":1022,"duration_ms":23135,"significance":"A practical, modular pipeline that balances identity concealment with attribute preservation would be useful for releasing diverse urban AV datasets under privacy constraints (GDPR-style rules, cultural cases such as veils). The explicit comparison of Roop versus Ghost-v2 on real street frames, the inclusion of restoration before swapping, and the focus on gaze/expression proxies are concrete engineering contributions. Credit is due for demonstrating multi-pedestrian automation and for highlighting failure modes of head-swapping on veiled subjects. Significance remains limited until downstream utility is measured beyond proxies.","major_comments":[{"comment":"The central usability claim for pedestrian-intention and trajectory-prediction models is supported only by four proxy metrics on high-quality close-ups (Table II) and a binary looking/not-looking extractor on a handful of clearly visible pedestrians (Section III-E, Fig. 5). No intention or trajectory model is retrained or evaluated on the anonymized frames, so the leap from residual landmark/blendshape/gaze differences (and from CodeFormer resize-restore) to retained training utility is untested. Given the Introduction’s own statement that facial cues are the main features required by these models, an end-to-end experiment is load-bearing for the dual-purpose claim.","section":"Section III-E / Extra Phase and Table II"},{"comment":"Table II reports mean landmark difference, blendshape difference, identity cosine similarity and gaze cosine similarity without sample size N, variance/std, or any statistical test. The text only says “mean average of a sample of facial close-up images.” Without these quantities it is impossible to judge whether Roop’s reported advantages (e.g., blendshape 1.898 vs 2.048) are reliable or practically meaningful.","section":"Table II"},{"comment":"Quantitative metrics appear confined to high-quality close-ups; results on the actual target domain (Egyptian street frames, occlusions, veils, multi-scale pedestrians) are almost entirely qualitative (Tables III–IV, Figs. 3–4). A quantitative evaluation of the same four metrics (or of a downstream model) on a representative subset of Egy-DRiVeS is needed to substantiate the claim that the pipeline is “tailored to satisfy the privacy needs of the Egy-DRiVeS dataset.”","section":"Section III-B–D and Tables III–IV"}],"minor_comments":[{"comment":"The text refers to “Table 5” for the looking/not-looking experiment while the caption and figure list show Fig. 5; numbering should be consistent.","section":"Section III-E"},{"comment":"Source-face selection is described as “fixed and randomly chosen” with no disclosure of the number of sources, demographic matching criteria, or whether the same source is reused across frames; a short paragraph or appendix would improve reproducibility.","section":"Section II-A"},{"comment":"Inference-time remark (“3-minute inference time”) appears only in the conclusion; reporting per-stage or per-frame timings in the results section would help readers assess video applicability.","section":"Section IV"},{"comment":"Several citations (Roop GitHub, Ghost-v2 arXiv) are preprints or repositories; a brief note on version/commit used would aid exact reproduction.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The pipeline is architecturally close to the 3PFS framework the authors themselves cite; the novelty delta (mainly Roop substitution + CodeFormer + Egyptian-specific handling) should be stated more sharply for the editor. Evaluation depth is currently light for a top-tier CV/ITS venue, but the engineering contribution is real and fixable with the requested end-to-end and statistical additions. Scope is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The punchline is simple: this is a clean engineering pipeline (YOLOv11 + SCRFD + CodeFormer + Roop + SeamlessClone) built for the real mess of Egy-DRiVeS street frames, and the head-to-head with Ghost-v2 is the useful part. Roop keeps head shape, hair, and expressions better and does not invent ghostly hair on veiled women; Ghost-v2 does. That comparison is new relative to the 3PFS paper and is worth having on the record.\n\nWhat the paper does well is the modular staging and the qualitative failure modes. Distant faces, occlusion, and veils are shown honestly; the looking/not-looking check on a handful of clear pedestrians is a sensible first sanity test. The four proxy metrics (landmarks, blendshapes, identity cosine, gaze cosine) are external and not circular. Citations to DeepPrivacy, CIAGAN, Ghost-v2, and 3PFS are appropriate; the authors do not over-claim a new learning principle.\n\nThe soft spot is exactly the one the stress-test flags, and it is real but not fatal for an engineering note. Usability for intention and trajectory models is asserted from those proxies plus one binary feature extractor. No intention model is retrained, no error bars or sample sizes appear on the metric tables, and CodeFormer’s resize-restore step is never measured for downstream damage. Source-face choice is also free and unanalyzed. Those gaps keep the central claim conditional rather than demonstrated.\n\nThis is for people who actually ship privacy filters on regional AV datasets and need a working recipe that handles veils. It is not for theorists looking for a new anonymization objective. I would send it to peer review; a referee can demand the missing end-to-end numbers and artifact release without the paper collapsing. Worth a look if you work on ITS data release; skip if you only care about new face-swap architectures.","headline":"Practical five-stage assembly that shows Roop beats Ghost-v2 on Egyptian street cases (veils, occlusion), but the AV-usability claim still rests only on proxies.","tokens_in":9713,"tokens_out":499,"would_cite":false,"duration_ms":10919,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A five-stage face-swapping pipeline conceals pedestrian identities in AV street datasets while keeping facial expressions, pose, and gaze usable for intention models.","keywords":["pedestrian privacy","face swapping","data usability","autonomous vehicles","ITS datasets","facial attribute preservation","Roop","Egy-DRiVeS"],"falsifier":"Retrain a standard pedestrian-intention or trajectory-prediction model on the anonymized Egy-DRiVeS frames and measure a clear drop in accuracy relative to the same model trained on the original frames.","tokens_in":9684,"feed_emoji":"🚗","tokens_out":584,"duration_ms":5333,"temperature":0.7,"pith_summary":"Autonomous-vehicle training needs large street datasets of real pedestrians, but those images expose identities and create privacy risks. Simple blurring destroys the facial cues that intention and trajectory models rely on. This paper builds a five-stage pipeline that first finds pedestrians and faces, restores low-resolution faces, swaps each face with a fixed public source using the Roop model, then blends the result back into the frame. The goal is dual: hide who the person is while leaving expressions, head pose, and eye-gaze direction intact enough that downstream looking/not-looking and intention predictors still work. On Egyptian street imagery that includes veils and occlusions, Roop is shown to beat Ghost-v2 on realism and attribute preservation, and a simple gaze feature extractor continues to label looking versus not-looking correctly after the pipeline runs.","feed_headline":"Face-swap pipeline hides pedestrians, keeps gaze for AVs","feed_subtitle":"Five stages with Roop anonymize street images while intention-model cues survive.","key_machinery":"The five-stage pipeline (pedestrian detection, face detection, quality enhancement, face swapping with Roop, Poisson blending) that transfers only identity while holding expression, pose and gaze fixed.","core_discovery":"A five-stage pipeline built around Roop face swapping, after YOLOv11 pedestrian detection, SCRFD face detection and CodeFormer restoration, conceals pedestrian identity on Egy-DRiVeS-style street images while preserving facial expressions, head pose and eye-gaze direction sufficiently for usability in subsequent pedestrian-intention tasks, and Roop outperforms Ghost-v2 on the reported metrics and on challenging cases such as occlusion and veiled faces.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Roop face-swap pipeline hides IDs, keeps gaze for AV training","Five-stage Roop swap anonymizes pedestrians, retains pose cues","Face swap with Roop shields privacy, preserves intention features","Roop outperforms Ghost-v2 in dual-purpose privacy pipeline","Pipeline swaps faces via Roop: IDs gone, eye-gaze stays for ITS"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That four proxy metrics plus a single looking/not-looking extractor on a handful of clearly visible pedestrians are enough to guarantee that full intention and trajectory models will keep their training value after anonymization.","fun_headline_variants_meta":{"raw":{"variants":["Roop face-swap pipeline hides IDs, keeps gaze for AV training","Five-stage Roop swap anonymizes pedestrians, retains pose cues","Face swap with Roop shields privacy, preserves intention features","Roop outperforms Ghost-v2 in dual-purpose privacy pipeline","Pipeline swaps faces via Roop: IDs gone, eye-gaze stays for ITS"]},"model":"grok-4.5","effort":"low","cost_usd":0.004392,"raw_usage":{"total_tokens":1278,"prompt_tokens":771,"num_sources_used":0,"completion_tokens":99,"cost_in_usd_ticks":43920000,"prompt_tokens_details":{"text_tokens":771,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":408,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":771,"tokens_out":99,"duration_ms":4893,"temperature":1.0,"reasoning_tokens":408,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T07:59:38.226412+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Retrain a standard pedestrian-intention or trajectory-prediction model on the anonymized Egy-DRiVeS frames and measure a clear drop in accuracy relative to the same model trained on the original frames.","supporting_citations":[],"review_version":1}