{"id":"81adad0d-d271-49bd-95f4-e795738c27d4","arxiv_id":"2504.20370","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ABO offloads raw Bayer-pattern camera frames through a tile-wise neural codec and adaptive bandwidth controller, improving AR DNN detection throughput, latency, and accuracy over RGB/JPEG offloading.","lead":"This paper proposes ABO, a system that sends raw camera sensor frames instead of fully processed color images to an edge server for augmented reality object detection. It reports that this approach increases frame throughput by 40%, cuts end-to-end latency by 30%, and improves detection accuracy by up to 15% compared to existing offloading systems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ABO's reported accuracy gain is confounded with task-aware distillation: Table 1 shows ABO-noDistill (RAW, no KD) is less accurate than JPEG, so abandoning Bayer alone does not explain the claimed 15% improvement.","rationale":"The reader identified the narrow self-collected dataset and missing train/test separation as the weakest assumption. While valid, the more load-bearing issue is internal to the evidence: the paper's headline accuracy gain is driven by the knowledge-distillation objective, not by the RAW format itself. Table 1 directly shows that the RAW codec without distillation is worse than JPEG, so the 'abandon Bayer' factor alone cannot explain the reported accuracy improvement. The system-level throughput and latency gains are plausible and well-supported by the pipelining and smaller single-channel frames, but the paper's central causal attribution is not isolated. A straightforward RGB-input ablation would settle whether the novelty is in the RAW format or in the task-aware codec/pipeline; without it, the paper should only be accepted conditionally, requiring this experiment plus data/code release. This is why I keep the conditional verdict but with a more specific condition than the reader's generalizability concern.","tokens_in":23452,"tokens_out":4965,"duration_ms":53369,"concrete_test":"Train an RGB-input twin of the ABO codec: same encoder/decoder architecture with 3 input channels, same tile partitioning, same knowledge-distillation loss from YOLOv5, and same dynamic controller; run it on the same five test segments and bandwidth traces. Compare mAP, average frame size, end-to-end latency, and FPS against ABO-RAW and JPEG. If RGB-ABO-Distill reaches within 1% mAP of ABO-RAW at similar frame sizes, the accuracy claim is not attributable to abandoning the Bayer filter, and the paper should be revised to frame the contribution as task-aware neural compression plus pipelining. If ABO-RAW remains significantly more accurate or smaller, the RAW-specific claim is supported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that skipping demosaicing improves end-to-end DNN offloading is tested only through a confounded comparison. In Table 1 (Section 2.2), the RAW neural codec without knowledge distillation (ABO-noDistill) yields F1=0.873, mAP=0.878, below JPEG's 0.883/0.899. The accuracy advantage of ABO (0.923/0.937) appears only after the distillation objective of Section 3.2.4 is applied. Thus the 'up to 15% accuracy improvement' in the Abstract is caused by task-aware codec training, not by the decision to operate on pre-demosaic RAW data. No experiment in the paper holds the codec and pipeline fixed while varying only the input format; the ablation in Section 5.7 removes distillation and tile selection, but never an RGB-input variant. Consequently, the paper cannot support the title's causal claim that abandoning the Bayer filter is what delivers the accuracy benefit, though the throughput/latency savings from 1-channel RAW and pipelining remain plausible.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ABO, an adaptive offloading framework for AR that compresses pre-demosaic RAW frames with a tile-wise neural codec and a dynamic controller that selects tiles and codec configurations based on estimated bandwidth. The central claim is that skipping demosaicing for offloaded DNN analytics improves end-to-end throughput, latency, and downstream task accuracy relative to standard RGB/JPEG offloading, and the authors support this with a hardware prototype, a self-collected RAW dataset, several baselines, and ablations. The claimed gains are 40% more throughput, 30% less latency, and up to 15% better DNN accuracy than state-of-the-art baselines.","tokens_in":23661,"tokens_out":6135,"duration_ms":60345,"significance":"If validated, the idea of decoupling demosaicing from offloading is a timely and practically relevant contribution to edge-assisted AR, and the throughput/latency savings from 1-channel RAW transmission and pipelining are plausible. The paper also contributes a hardware prototype, a new RAW dataset, and a careful ablation of the controller components. However, the headline accuracy benefit is confounded with task-aware codec training, and the evaluation uses a self-collected dataset without an explicit train/test split, so the causal claim about abandoning the Bayer filter is not yet established. The paper does not release code or data, which limits reproducibility despite its otherwise thorough systems evaluation.","major_comments":[{"comment":"The accuracy advantage attributed to RAW offloading is confounded with task-aware distillation. Table 1 shows that the RAW neural codec without knowledge distillation (ABO-noDistill, F1=0.873, mAP=0.878) is less accurate than JPEG (F1=0.883, mAP=0.899), and the advantage of ABO-Distill (F1=0.923, mAP=0.937) appears only after adding the distillation loss of Eq. (2). Since no experiment varies only the input format while keeping the codec and training objective fixed, the title's causal claim that abandoning the Bayer filter improves accuracy is not supported. Please add an RGB-input variant of the same neural codec trained with the same distillation objective, or a demosaiced-RAW variant of ABO, to isolate the effect of the input format from the effect of task-aware codec training.","section":"Section 2.2, Table 1; Section 3.2.4"},{"comment":"The evaluation uses a self-collected RAW dataset and five consecutive-frame segments, but no explicit train/test split or cross-validation is described. It is unclear whether the five test segments, the frames used to train the codec and the YOLO model, and the 'small profiling data set' used for the LUT in Section 3.3.2 are disjoint. Without this information, the Pareto curves in Figure 10 and the end-to-end results in Figures 11-13 could be optimistically biased by overfitting to one sensor and one scene distribution. Please specify the exact split used for training, profiling, and testing, and, if possible, validate on a second sensor or an independent scene set.","section":"Section 5.1.1 and Appendix D"},{"comment":"The claimed 'up to 15%' accuracy improvement in the Abstract is inconsistent with Table 1, where ABO-Distill improves over JPEG by 0.040 F1 and 0.038 mAP (about 4-5% relative). Section 5.3 states that ABO has 'over 15-20% better task accuracy' without defining whether the comparison is absolute or relative, and without identifying the specific operating point on the Pareto curve. Please reconcile these numbers, state the bandwidth at which the 15% figure is achieved, and report variance or confidence intervals for the accuracy measurements.","section":"Abstract and Section 5.3"}],"minor_comments":[{"comment":"The text contains the typo 'Appandix C'; it should be 'Appendix C'.","section":"Section 3.2.1"},{"comment":"The text refers to a 'power moniter'; this should be 'power monitor'.","section":"Section 5.8"},{"comment":"The caption contains 'original RA W image' with a stray space; please correct to 'RAW'.","section":"Figure 3 caption"},{"comment":"The limitation statement about limited RAW image data is valuable, but it appears only in the appendix; consider summarizing it in Section 5 so the evaluation's scope is transparent to readers of the main body.","section":"Appendix G"},{"comment":"The paper does not state whether code and data will be released; given the self-collected dataset and prototype, a release statement would substantially aid reproducibility.","section":"Section 5.1.1 and Implementation"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid systems contribution with a credible prototype, but the central accuracy claim needs the isolation experiment described in the major comments before I can support acceptance. I would also encourage the editors to require a clear data/code availability statement, consistent with the reproducibility standards expected for systems papers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"ABO is a well-built systems paper whose central accuracy claim doesn't survive contact with its own Table 1. The paper says skipping demosaicing improves DNN accuracy by up to 15%, but Table 1 shows the RAW codec without knowledge distillation (ABO-noDistill) at 0.873 F1 / 0.878 mAP, below JPEG's 0.883/0.899. The improvement appears only after adding the distillation loss in Section 3.2.4. So the accuracy gain is a property of task-aware codec training, not of abandoning the Bayer filter. No experiment holds the pipeline fixed and varies only the input format, so the title's causal story is unsupported.\n\nWhat is genuinely new: ABO is the first offloading framework I know that operates on pre-demosaic RAW frames for DNN analytics, with a configurable tile-wise RAW codec and a controller that picks tiles and codec settings from bandwidth. The hardware prototype (IMX178 + Jetson Nano + RTX 4090) is real, the baselines are appropriate (JPEG, DeepCOD, PNC, Reducto), and the ablation isolates distillation and tile selection. The throughput and latency gains—40% throughput, 30% latency—come from single-channel RAW plus pipelining that hides demosaicing. Those claims are plausible and are the paper's real contribution.\n\nSoft spots, in order. First, the accuracy claim is overstated and confounded; the paper should either re-run the comparison with a RAW codec trained without distillation versus an RGB codec trained without distillation, or restrict the claim to task-aware codec training. Second, the dataset is self-collected with no described train/test split, and neither data nor code is released, so the numbers can't be independently checked. That matters more for a systems paper whose results rest on one sensor and five video segments. Third, the memory footprint (1080 MB) is worth a sentence on why that's acceptable for AR glasses, though it's a minor point.\n\nThe authors are honest in Appendix G about limited RAW data, so I'd call the thinking serious. The stress-test concern stands: the abstract's 'up to 15% accuracy improvement' is a distillation effect, not a Bayer-filter effect. With a controlled experiment and released data, this could be a decent systems contribution for the mobile AR and edge video analytics crowd. As it stands, I'd send it to peer review but insist on the confounding experiment before acceptance.","headline":"Solid RAW offloading system whose headline accuracy gain is really a knowledge-distillation effect; the claimed Bayer-filter benefit is unsupported by the paper's own Table 1.","tokens_in":24201,"tokens_out":2214,"would_cite":false,"duration_ms":22203,"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":"ABO shows that AR systems can skip demosaicing for offloaded DNN analytics, transmitting compressed RAW tiles instead of RGB frames, and reports 40% higher throughput, 30% lower latency, and up to 15% better DNN accuracy than…","keywords":["augmented reality","DNN offloading","RAW image compression","neural codec","Bayer filter","demosaicing","adaptive transmission","edge computing"],"falsifier":"Run ABO against the same baselines on RAW frames from a different sensor (different Bayer pattern, bit depth, or form factor) with a held-out test set from multiple environments; if the mAP gain over JPEG drops below the reported margin, or demosaicing no longer dominates the client-side latency, the central claim fails to generalize.","tokens_in":23246,"feed_emoji":"📡","tokens_out":5905,"duration_ms":57154,"temperature":0.7,"pith_summary":"ABO claims the conventional color image pipeline works against augmented reality: demosaicing RAW sensor frames into RGB triples the data sent to an edge server for DNN analysis and adds tens of milliseconds of on-device computation, yet does not improve detection accuracy. The paper proposes sending single-channel RAW frames through a lightweight tile-wise neural codec, transmitting only tiles that contained objects in the previous frame, and adapting the codec configuration to current bandwidth, while demosaicing the local rendering stream in parallel. On a prototype AR goggle, the system reports 40% more frame-processing throughput, 30% lower end-to-end latency, and up to 15% higher downstream DNN accuracy than JPEG and three state-of-the-art offloading baselines, and it is the only tested system that reaches the real-time targets of under 40 ms latency and over 25 FPS. If these results generalize, AR devices can get responsive, accurate perception without faster radios or stronger on-board GPUs.","feed_headline":"Skipping demosaicing speeds up AR analytics by 40 percent","feed_subtitle":"A RAW-tile neural codec hides color interpolation behind network transmission and beats JPEG pipelines.","key_machinery":"The load-bearing object is a configurable tile-wise RAW neural codec built as an asymmetric autoencoder. A single convolution layer on the device compresses each selected tile into a feature map; on the server, a deep residual decoder reconstructs the RAW tile. Four configurations are obtained by varying convolution stride (2 or 4) and output channels (4 or 8), and a pluggable decoder head lets one shared decoder serve all four. The RAW frame is disassembled into four CFA color channels before encoding to preserve color information, then reassembled after decoding. Around this codec, a dynamic transmission controller picks tiles by overlap with the previous frame's detection boxes and picks the profiled configuration whose bandwidth cost fits the estimated available bandwidth. Training includes knowledge distillation from the downstream object detector, so the codec learns to keep task-relevant structure rather than only pixel fidelity.","core_discovery":"The paper's central discovery is that for machine-centric offloading, the single-channel Bayer RAW frame is the transmitted artifact of choice, not the demosaiced RGB image. ABO shows that a task-aware neural codec can compress RAW tiles to sizes comparable to JPEG while preserving or improving the information the downstream detector needs; knowledge distillation from the detector even sharpens object-relevant edges. Because the RAW stream has one third the channel count and the encoder is a single convolution layer, the client spends little time compressing, and the 25–45 ms demosaicing step can run in parallel with transmission and server inference rather than before them. The resulting pipeline reaches 29 FPS and 37 ms end-to-end latency on an embedded-class client over a 200 Mbps link, and the controller's tile selection plus configuration calibration keeps accuracy high when bandwidth varies. The paper states that among all tested systems, ABO is the only one that met the real-time experience target.","pith_inferences":["If the RAW-input advantage holds across sensors, camera ISPs could expose a bypass mode that outputs compressed RAW tiles for machine vision, saving sensor power and silicon area currently spent on demosaicing.","The same controller pattern—profiled look-up table, content-aware tile selection, lightweight bandwidth estimate—could extend to other raw sensor streams or to multi-modal AR (audio, depth), though the paper only evaluates image object detection.","A quicker test of generality would be to synthesize RAW frames from a public RGB detection benchmark with a different CFA pattern and camera response function; if ABO's 15% accuracy margin does not survive, the gain may be partly dataset-specific.","The paper's own limitation notes point to tile-level codec calibration and large-scale RGB pretraining as natural next steps, both of which could reduce bandwidth further and loosen the reliance on large RAW collections."],"forward_implications":["AR offloading can hit the 40 ms and 25 FPS real-time bar on current embedded hardware and a 200 Mbps link, without upgrading the radio or the on-board processor.","Machine-centric image codecs trained with downstream task loss dominate human-view codecs like JPEG on the accuracy-bandwidth frontier.","Demosaicing can be moved out of the critical path entirely: rendering still gets a full-color frame, but the offloaded analysis never waits for color interpolation.","Bandwidth fluctuations can be absorbed by adapting tile selection and codec configuration per frame, keeping task accuracy near the Pareto-optimal frontier.","RAW-based offloading degrades less in dim light and fast motion than JPEG-based offloading, which matters when AR users move and rooms are poorly lit."],"supporting_citations":[{"why":"Supplies the standard JPEG compression baseline that ABO must beat in frame size, accuracy, latency, and throughput.","marker":"[53]"},{"why":"DeepCOD is the neural offloading baseline whose frame-level encoding ABO compares against.","marker":"[61]"},{"why":"PNC provides the progressive neural compression baseline using stochastic tail-drop.","marker":"[55]"},{"why":"Reducto is the frame-selection offloading baseline that ABO compares against on bandwidth efficiency.","marker":"[28]"},{"why":"Knowledge distillation from the downstream DNN is the training objective that gives ABO its task-aware edge enhancement.","marker":"[19]"},{"why":"YOLO is the downstream object detector whose loss drives codec training and whose mAP and F1 measure task accuracy.","marker":"[45]"},{"why":"The IMX178 sensor in the hardware prototype produces the RAW frames that define the evaluation domain.","marker":"[18]"},{"why":"Residual blocks form the deep decoder architecture that reconstructs transmitted tiles on the server.","marker":"[16]"},{"why":"Real-world and generated bandwidth traces are replayed to evaluate ABO's adaptation under network dynamics.","marker":"[22]"}],"fun_headline_variants":["Raw Bayer frames win for AR offload analytics","Skip demosaicing: ABO cuts AR latency by 30%","ABO: adaptive RAW offloading boosts AR throughput 40%","Bayer direct: ABO's RAW codec beats JPEG pipelines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the self-collected RAW dataset and five consecutive-frame segments, captured with one prototype sensor in campus office scenes, are representative enough that the reported 40% throughput, 30% latency, and 15% accuracy gains transfer to other AR hardware, scenes, lighting, and DNN tasks.","fun_headline_variants_meta":{"raw":{"variants":["Raw Bayer frames win for AR offload analytics","Skip demosaicing: ABO cuts AR latency by 30%","ABO: adaptive RAW offloading boosts AR throughput 40%","Bayer direct: ABO's RAW codec beats JPEG pipelines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1319,"prompt_tokens":980,"completion_tokens":339,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":267}},"tokens_in":596,"tokens_out":339,"duration_ms":4091,"temperature":1.0,"reasoning_tokens":267,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:30:32.588520+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ABO against the same baselines on RAW frames from a different sensor (different Bayer pattern, bit depth, or form factor) with a held-out test set from multiple environments; if the mAP gain over JPEG drops below the reported margin, or demosaicing no longer dominates the client-side latency, the central claim fails to generalize.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DeepCOD is the neural offloading baseline whose frame-level encoding ABO compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PNC provides the progressive neural compression baseline using stochastic tail-drop."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Reducto is the frame-selection offloading baseline that ABO compares against on bandwidth efficiency."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The IMX178 sensor in the hardware prototype produces the RAW frames that define the evaluation domain."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Real-world and generated bandwidth traces are replayed to evaluate ABO's adaptation under network dynamics."}],"review_version":1}