{"id":"2d66b2c3-4795-4b2c-a2f7-d3ec8d77f104","arxiv_id":"2603.11917","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A 1.3M-parameter CNN with ROI-implicit prompting and SAM3 distillation reaches ~65% mIoU on COCO/LVIS and 11.82 ms INT8 inference fully in-sensor on the Sony IMX500.","lead":"PicoSAM3 is a 1.3M-parameter CNN that does promptable object segmentation inside a camera chip (Sony IMX500) in about 12 ms. It shows that ROI-cropped distillation from large SAM models can make high-quality masks practical on extreme-edge sensors for glasses and IoT.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"Table I’s “beats SAM-H” numbers are not protocol-matched; the central superiority claim rests on that mismatch.","rationale":"The reader correctly isolates the single load-bearing soft spot: protocol mismatch between Table I baselines and the 96×96 ROI evaluation that defines PicoSAM3. Everything else (architecture, distillation ablations in Table II, INT8 size, IMX500 11.82 ms latency, public code) is internally consistent and supports the feasibility half of the claim. The comparative half does not, until the baselines are re-measured under the same crop protocol the paper itself admits collapses large SAMs. That single check either rehabilitates or forces retraction of the “surpasses SAM-H / outperforming existing … baselines” language; no other concern is as central. Verdict therefore stays CONDITIONAL, matching the reader.","tokens_in":14960,"tokens_out":572,"duration_ms":5182,"concrete_test":"Re-evaluate SAM-H, FastSAM, TinySAM, EdgeSAM, and LiteSAM on the exact same COCO/LVIS ROI-crop pipeline used for PicoSAM3 (10% padded square crop → 96×96, box prompt to the teacher/decoder where applicable). Replace the Table I mIoU columns with these same-protocol numbers. If PicoSAM3’s advantage over SAM-H falls below ~3–5 points or vanishes, the superiority rhetoric must be withdrawn; if it remains large, the claim is protocol-solid.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper’s strongest comparative claim (Abstract; §V.A; Table I) is that PicoSAM3 “surpasses SAM-H” by +11.85 mIoU on COCO (65.45% vs 53.6%) and similarly beats FastSAM/TinySAM/EdgeSAM/LiteSAM. Those baseline numbers are the published full-image / high-resolution figures. In §V.B and Fig. 3 the authors themselves show that large SAMs suffer severe feature collapse on the exact 96×96 ROI-crop protocol used for PicoSAM3 (~25% mAP for SAM2.1 Large). Table I therefore mixes two incompatible evaluation regimes: PicoSAM3 is scored only on centered 96×96 crops with implicit box prompts, while the foundation and most edge baselines are listed at their original published scales. If the headline gains disappear (or reverse) once every baseline is re-run under the identical crop + resize + box-prompt protocol, the “outperforming … at similar or lower complexity” narrative no longer supports the central claim of high-quality promptable segmentation relative to prior art—only the narrower (still valuable) claim of feasible in-sensor deployment.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"PicoSAM3 is a 1.3–1.4 M-parameter fully convolutional, ROI-prompted segmentation student designed for the Sony IMX500’s <8 MB SRAM and restricted operator set. Prompts are encoded implicitly by centered padded square crops (no extra prompt channels); the architecture extends PicoSAM2 with a dilated bottleneck, ECA, and a refinement head; training uses adaptive distillation from SAM2/SAM3 soft masks plus ground-truth and area-preservation losses. On COCO/LVIS ROI-crop evaluation the FP32 model reports 65.45%/64.01% mIoU; INT8 PTQ yields 1.31 MB and 11.82 ms end-to-end latency on the IMX500 with <0.2% mIoU drop. Ablations (Table II) attribute large gains to ROI cropping and SAM3 distillation over supervised training.","tokens_in":15266,"tokens_out":1292,"duration_ms":15297,"significance":"If the in-sensor results hold under clarified evaluation, the paper is a concrete systems contribution for IEEE Sensors Journal: it shows that spatially flexible, box-style promptable segmentation can run fully on-sensor at ~84 FPS under real memory/operator constraints, with public code and measured silicon latency. Strengths include clean ablations separating ROI, teacher, architecture, and quantization; near-lossless INT8; and a hardware-aligned implicit-prompt design that maps to the IMX500 ROI API. The work is valuable even if headline comparisons to full-resolution foundation models are revised, because the deployment and PicoSAM2-to-PicoSAM3 progression are independently useful.","major_comments":[{"comment":"Table I / Abstract / §V.A: The central comparative claim that PicoSAM3 “surpasses SAM-H” by +11.85 mIoU on COCO (65.45% vs 53.6%) and similarly beats FastSAM/TinySAM/EdgeSAM/LiteSAM mixes incompatible protocols. PicoSAM3 is scored only on centered 96×96 ROI crops with implicit box prompts; the listed foundation/edge numbers are published full-image, high-resolution figures. §V.B and Fig. 3 themselves show large SAMs suffer severe feature collapse under the 96×96 crop regime (~25% mAP for SAM2.1 Large). Without a same-protocol re-evaluation column (or a clear split of “published full-res” vs “96×96 ROI-crop” scores), the “outperforming … at similar or lower complexity” narrative does not support superiority over prior art—only the narrower (still important) claim of feasible in-sensor deployment and gains over PicoSAM2 under matched conditions.","section":null},{"comment":"§IV.A–B and evaluation protocol: “Promptable” is used throughout, but the system supports only single-object, box-like conditioning via hardware ROI crop (no point prompts, no multi-mask decoding, no iterative refinement at inference beyond re-cropping). Table I labels PicoSAM3 “Box” while several baselines are full SAM-style promptable models. Please define the supported prompt interface explicitly, report whether multi-object or point-prompt settings are out of scope, and ensure baseline comparisons use the same prompt type and crop protocol where claims of outperformance are made.","section":null},{"comment":"§IV.D / Table I: Training is reported as one epoch on COCO with batch 64. For a distillation student claiming large gains over supervised training and SOTA among edge models, one epoch is unusually short and risks under-training or teacher-cache memorization. Provide training curves, multi-epoch ablations, or validation that longer training does not change the ranking vs PicoSAM2 and vs same-protocol baselines; otherwise the absolute mIoU numbers and distillation gains (+14.5% in the abstract) need stronger evidence of stability.","section":null}],"minor_comments":[{"comment":"§V: “PyTorch 3.7” is not a released version; correct to the actual version used.","section":null},{"comment":"References: Milletari et al. V-Net appears twice ([49] and [50]); deduplicate.","section":null},{"comment":"Table I: Q-PicoSAM3 size is listed as 1.30 MB in the table and 1.31 MB in the text; unify.","section":null},{"comment":"Fig. 3 caption and §V.B: Clarify that “feature collapse” of large SAMs is under the authors’ 96×96 crop protocol so readers do not misread the Pareto plot as a general ranking of foundation models.","section":null},{"comment":"Abstract vs body: Abstract says “1.3M parameters” and “distillation from SAM2 and SAM3”; body uses 1.37M and primarily reports SAM3 for the final model—align wording.","section":null},{"comment":"Eq. (6)–(9): Temperature-scaled sigmoid is written σ_τ(x)=σ(τ·x); confirm whether τ multiplies logits before or after scaling as intended for “sharper” distributions (τ=5 typically softens if used as divisor).","section":null},{"comment":"§III / Fig. 1: Minor grammar (“runs fully in-sensors”); polish for camera-ready.","section":null}],"recommendation":"major_revision","confidential_remarks":"Fit for IEEE Sensors Journal is good if the authors reframe Table I around matched-protocol and on-sensor metrics rather than beating published SAM-H full-res scores. The silicon result and PicoSAM2→3 progression are the real contribution; the protocol mismatch is fixable without new hardware work but is currently load-bearing for the abstract’s superiority language. I would not reject on novelty relative to PicoSAM2 alone—the ROI flexibility, SAM3 distillation, and measured 11.82 ms are incremental but legitimate for this venue."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful result here is systems, not a new theory of segmentation: a 1.37M CNN student with implicit ROI cropping, ECA + dilated bottleneck, SAM3 distillation, and a working INT8 path that hits 11.82 ms end-to-end on the Sony IMX500 under its real SRAM/operator limits. That is a clean extension of their PicoSAM2 line—flexible box-style ROI instead of centered-point, better teacher, small architecture tweaks—and they ship code.\n\nWhat they do well is the hardware-aware loop. Ablations in Table II separate ROI cropping, teacher (SAM2 vs SAM3), architecture, and quantization; the INT8 drop is tiny (~0.11 mIoU). Distillation vs supervised is large and believable. The deployment story (crop via sensor ROI API, 96×96, no exotic ops) is concrete and matches the claim that promptable masks can run at the sensor without cloud offload. For smart-glasses / privacy IoT people, that is the part that matters.\n\nThe soft spot is real but bounded. Table I and the abstract lean hard on “+11.85 mIoU over SAM-H” and similar wins vs FastSAM/TinySAM/etc. Those baselines are the published full-image / high-res numbers. Later (§V.B, Fig. 3) the authors themselves show large SAMs collapse on the exact 96×96 crop protocol (~25% mAP). So the headline superiority narrative mixes incompatible regimes. Once you strip that rhetoric, what remains is still solid: among models that actually fit the IMX500, they improve on their own prior work and demonstrate feasibility. The free parameters (padding, τ, ρ, channels) are ordinary; nothing looks circular.\n\nWho it is for: extreme-edge vision and in-sensor people, not general SAM theorists. A serious referee should take it, demand a same-protocol baseline column (or a clear disclaimer that foundation numbers are not apples-to-apples), and keep the hardware result. I would cite the deployment and the ROI-implicit recipe if I were working on sensor-side masks; I would not cite the SAM-H comparison without the fix. Worth engaging.","headline":"Solid IMX500 systems paper with real in-sensor ROI masks; Table I’s “beats SAM-H” numbers are protocol-mismatched and should not be taken at face value.","tokens_in":15948,"tokens_out":549,"would_cite":true,"duration_ms":5137,"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 1.3-million-parameter CNN can deliver high-quality promptable object segmentation inside a vision sensor in real time.","keywords":["Edge AI","In-sensor computing","Promptable segmentation","Knowledge distillation","Quantization","Real-time segmentation","Segment Anything Model","Smart sensors"],"falsifier":"Run SAM-H, FastSAM, TinySAM and EdgeSAM on exactly the same COCO and LVIS ROI crops resized to 96 by 96 that PicoSAM3 receives; if their mIoU then equals or exceeds PicoSAM3’s 65.45 percent / 64.01 percent, the superiority claim under in-sensor conditions fails.","tokens_in":15793,"feed_emoji":"📷","tokens_out":1023,"duration_ms":16509,"temperature":0.7,"pith_summary":"This paper shows that promptable visual segmentation need not stay on large GPUs or even on a phone CPU: it can run fully inside a stacked smart camera sensor. PicoSAM3 is a dense convolutional network of about 1.3 million parameters that treats a hardware region-of-interest crop as an implicit box prompt, adds light channel attention and a dilated bottleneck, and is trained by distilling soft masks from the large SAM2 and SAM3 teachers. On standard COCO and LVIS ROI-crop tests it reaches 65.45 percent and 64.01 percent mean intersection-over-union, while the INT8 version fits in 1.31 MB and finishes inference in 11.82 ms on the Sony IMX500. Ablations attribute large accuracy gains to the distillation step and to flexible ROI prompting rather than older centered-point prompts. The practical claim is that privacy-preserving, low-latency segmentation for smart glasses and IoT cameras is already feasible without leaving the sensor package.","feed_headline":"1.3M-param model segments objects in-sensor at 12 ms","feed_subtitle":"Tiny CNN distilled from SAM3 hits 65% mIoU on COCO while fitting the IMX500 memory limit","key_machinery":"Implicit ROI prompt encoding via centered object cropping (with padding) that is resized to 96 by 96, paired with offline knowledge distillation of temperature-scaled soft masks from SAM3 into a quantization-friendly CNN student. The crop itself supplies the spatial prior, so no extra prompt tensors are needed at inference and the pipeline maps directly onto the sensor’s hardware ROI mode.","core_discovery":"High-quality, spatially flexible promptable segmentation is feasible directly at the sensor. PicoSAM3 (1.37 M parameters) combines a U-Net-style dense CNN, dilated bottleneck, Efficient Channel Attention, ROI-crop implicit prompting, and distillation from SAM2/SAM3 to obtain 65.45 percent mIoU on COCO and 64.01 percent on LVIS; the INT8 model retains essentially the same accuracy at 1.31 MB and 11.82 ms end-to-end latency on the IMX500 while obeying its memory and operator limits.","pith_inferences":["The same ROI-crop plus distillation pattern could transfer to other stacked or always-on sensors that expose only RGB frames and a hardware crop API.","If large foundation models systematically collapse at 96-by-96 resolution, future edge work may need native low-resolution teachers rather than only compressing high-resolution ones.","Wearable AR pipelines could feed a tiny detector’s boxes straight into an in-sensor PicoSAM3-style mask head without ever leaving the camera package.","The adaptive weighting of teacher versus ground-truth loss by teacher confidence is a reusable recipe for distilling any large vision model onto extreme-edge students."],"forward_implications":["Promptable segmentation can execute fully in-sensor without cloud offload for latency- and privacy-sensitive devices.","Distillation from large SAM teachers can close most of the accuracy gap for sub-2 M-parameter CNNs operating on 96-by-96 ROI crops.","INT8 post-training quantization of this architecture incurs negligible mIoU loss while meeting extreme-edge memory budgets.","Flexible box/ROI prompting maps directly onto hardware ROI modes already present in intelligent vision sensors.","Hardware-aware dense CNNs avoid the unsupported operators and random memory access that block transformer SAM variants on the IMX500."],"fun_headline_variants":["PicoSAM3: 1.3M-param in-sensor segmentation at 11.8 ms","Edge CNN distilled from SAM hits 65% mIoU on IMX500","Real-time ROI masks inside the sensor with 1.3M params","Promptable segmentation on Sony IMX500 in under 12 ms","INT8 PicoSAM3 fits IMX500 for 11.82 ms on-device masks"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The claim that PicoSAM3 outperforms large foundation models rests on comparing numbers that may not have been measured under the same 96-by-96 ROI-crop protocol the tiny model actually uses.","fun_headline_variants_meta":{"raw":{"variants":["PicoSAM3: 1.3M-param in-sensor segmentation at 11.8 ms","Edge CNN distilled from SAM hits 65% mIoU on IMX500","Real-time ROI masks inside the sensor with 1.3M params","Promptable segmentation on Sony IMX500 in under 12 ms","INT8 PicoSAM3 fits IMX500 for 11.82 ms on-device masks"]},"model":"grok-4.5","effort":"low","cost_usd":0.006804,"raw_usage":{"total_tokens":1708,"prompt_tokens":813,"num_sources_used":0,"completion_tokens":113,"cost_in_usd_ticks":68040000,"prompt_tokens_details":{"text_tokens":813,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":782,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":813,"tokens_out":113,"duration_ms":6325,"temperature":1.0,"reasoning_tokens":782,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T22:31:37.101308+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run SAM-H, FastSAM, TinySAM and EdgeSAM on exactly the same COCO and LVIS ROI crops resized to 96 by 96 that PicoSAM3 receives; if their mIoU then equals or exceeds PicoSAM3’s 65.45 percent / 64.01 percent, the superiority claim under in-sensor conditions fails.","supporting_citations":[],"review_version":1}