{"id":"6a86a3e1-80ff-4c59-8a59-2e6aa3e1787e","arxiv_id":"2411.15678","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A new RAW-image detection benchmark and a RAW pre-training method with cross-domain distillation improve object detection under adverse conditions.","lead":"This paper introduces AODRaw, a dataset of 7,785 real camera RAW images with 135,601 labeled objects in 62 categories across nine lighting and weather conditions. It shows that pre-training object detectors directly on RAW images, assisted by an sRGB-trained teacher, improves detection in rain, fog, and low light without extra ISP modules.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RAW pre-training's 1.1% AP gain is confounded by the brightness/noise augmentation inserted during synthesis; without an sRGB pre-training control using the same augmentation, the improvement cannot be attributed to the RAW domain.","rationale":"The AODRaw dataset is a substantial contribution: 7,785 real RAW images, 135,601 instances, 62 categories, and 9 conditions, with detailed statistics and a benchmark of existing detectors. This part of the paper is solid and independently useful. The concern is specifically with the causal claim about RAW pre-training. The proposed method differs from the sRGB pre-training baseline not only in the input domain but also in the augmentation schedule: the unprocessing operation in Section 5.1 adds random brightness and noise, while the sRGB baseline uses the standard ImageNet pipeline. The paper's own robustness plots (Figures 5 and 6) demonstrate that sRGB pre-training is highly sensitive to such perturbations, making it plausible that the augmentation alone could produce the observed gains. Since the 1.1% AP improvement is the central evidence that RAW pre-training works, a controlled comparison is essential. This is not a fatal flaw; it can be resolved with one additional experiment, and the dataset and benchmark remain valuable regardless of the outcome. Therefore, conditional acceptance, as the reader recommended, is appropriate.","tokens_in":15961,"tokens_out":7244,"duration_ms":63391,"concrete_test":"Pre-train ConvNeXt-T on ImageNet-1K sRGB images with the same random brightness and shot-noise augmentation used for ImageNet-RAW (e.g., apply the unprocessing pipeline of [2] and then map the result back to sRGB with a fixed ISP), and fine-tune on AODRaw RAW under the same protocol as Table 3. If the resulting AP matches the 34.8% of RAW pre-training within 0.2%, the improvement is attributable to the augmentation schedule rather than the RAW domain; if it remains at or below the 33.7% sRGB baseline, the RAW-domain claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.1 describes synthesizing ImageNet-RAW by applying the unprocessing method [2] as a data augmentation, randomly adjusting brightness and noise each iteration. The key comparison in Section 5.2, RAW pre-training versus sRGB pre-training (34.8 vs 33.7 AP in Table 3 for Cascade RCNN/ConvNeXt-T), therefore differs in two variables: input domain (RAW vs sRGB) and augmentation schedule (random brightness/noise vs standard ImageNet). The paper does not include a control that pre-trains on sRGB ImageNet with the same brightness/noise augmentation. Figures 5 and 6 show sRGB pre-training degrades sharply under brightness/noise shifts (14.0% and 13.1% Top-1 drops), so an augmentation-only robustness gain is plausible and could account for part or all of the 1.1% AP. Additionally, fine-tuning applies gamma correction to real RAW (Section 4.1), while pre-training is performed on linear synthetic RAW from the unprocessing pipeline; this domain mismatch further muddies the definition of 'RAW pre-training'. Without the missing control, the central causal claim is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces AODRaw, a RAW-image object detection dataset with 7,785 high-resolution images, 135,601 annotated instances across 62 categories, and 9 combinations of light and weather conditions. It benchmarks multiple detection architectures under sRGB and RAW inputs, and proposes pre-training detectors on synthetic RAW images generated from ImageNet via the unprocessing method, augmented with cross-domain knowledge distillation from an off-the-shelf sRGB pre-trained teacher. The central claim is that RAW pre-training improves detection on real RAW images, especially under adverse conditions, without needing neural ISP adapters; the headline result is 34.8% AP for Cascade RCNN/ConvNeXt-T on RAW, versus 33.7% AP with sRGB pre-training.","tokens_in":16221,"tokens_out":5804,"duration_ms":50712,"significance":"The AODRaw dataset is a valuable new resource: it is substantially larger and more diverse than existing RAW detection datasets, with 62 categories and nine condition combinations, and the paper provides a careful statistical analysis of the data. The benchmark results across diverse backbones and detector families are extensive and internally consistent. The proposed pre-training recipe is practical and removes the need for learnable ISP modules at inference time. However, the paper's central causal claim—that the RAW input domain, rather than the accompanying brightness/noise augmentation or the distillation procedure, is responsible for the observed gains—is not yet established by the experiments. The dataset and benchmark alone would support a useful paper; the method claim requires additional controlled comparisons.","major_comments":[{"comment":"The comparison between RAW pre-training and sRGB pre-training is confounded by the data augmentation schedule. Section 5.1 states that in ImageNet-RAW synthesis, 'the unprocessed operation is inserted into the pipeline of data augmentations. Thus, we can randomly adjust the average brightness and simulated noise in each iteration.' The sRGB pre-training baseline in Table 3 (rows 'sRGB RAW') uses standard ImageNet augmentation without such brightness/noise randomization. Since Figures 5 and 6 show sRGB pre-trained models degrade by 14.0% and 13.1% under these shifts, a robustness gain from the augmentation alone is plausible. The paper needs a control that pre-trains on sRGB ImageNet with the same random brightness/noise augmentation (e.g., applied to the sRGB images) to attribute the reported 1.1% AP improvement to the RAW domain rather than to the augmentation schedule.","section":"Section 5.1, 5.2; Table 3"},{"comment":"The headline result is also confounded by the cross-domain distillation. The 'RAW RAW' rows in Table 3 correspond to the full proposed method that includes distillation, while the 'sRGB RAW' rows use plain sRGB pre-training without distillation. Table 8 shows that distillation contributes 0.5% AP (34.1% without distillation vs. 34.8% with distillation) within RAW pre-training. To isolate the effect of the RAW domain, the paper should report: (a) RAW pre-training without distillation versus sRGB pre-training with the same augmentation, and (b) sRGB pre-training with the same augmentation and with an analogous distillation from the same teacher. Without these controls, the 1.1% AP gain cannot be attributed to the RAW domain.","section":"Section 5.2; Table 8"},{"comment":"There is a mismatch between the pre-training and fine-tuning input domains. Section 4.1 says that real RAW images are 'further processed through gamma correction' before fine-tuning, while Section 5.1 describes synthesizing ImageNet-RAW using the unprocessing method, which produces linear RAW images with simulated noise. The paper does not state whether the synthetic pre-training data is gamma-corrected or linear. If the pre-training uses linear RAW and the fine-tuning uses gamma-corrected RAW, then the two domains differ by more than just the RAW-vs-sRGB distinction, and the term 'RAW pre-training' is ambiguous. The authors should either apply the same gamma correction to the synthetic pre-training data or explicitly justify the mismatch and demonstrate that it does not affect the main comparison.","section":"Section 4.1 vs. Section 5.1"},{"comment":"The dataset annotation section provides no information about annotator training, quality control, or inter-annotator agreement. Since AODRaw is a core contribution and is used to evaluate the pre-training method, the reliability of the annotations is essential. The authors should report the annotation protocol, the number of annotators, and a quality metric such as inter-annotator agreement or a manual verification rate on a subset, so that the benchmark results can be trusted as ground truth.","section":"Section 3.1"}],"minor_comments":[{"comment":"The paper contains a recurring typo where 'RAW' appears as 'RA W' (e.g., in the title and several headings); this should be fixed.","section":"Throughout"},{"comment":"The text says 'when reducing the brightness from 791 to 80', but the caption of Figure 5 states the maximum average brightness is 216. These numbers are inconsistent; please clarify the brightness scale and correct the text or figure.","section":"Section 5.1; Figures 5 and 6"},{"comment":"The 'Baseline' row reports 33.4% AP for ConvNeXt-T with sRGB pre-training and RAW fine-tuning, while the analogous row in Table 3 (Cascade RCNN/ConvNeXt-T, 'sRGB RAW') reports 33.7% AP. The discrepancy should be explained, since the reader cannot tell whether the baseline in Table 5 uses a different training recipe or a different definition.","section":"Table 5"},{"comment":"The text says 'too tiny objects with an area of less than 32^2 are ignored because they will disappear after down-sampling,' but the object-area thresholds for down-sampling are defined as [0,128^2), [128^2,320^2), and [320^2,+∞). The 32^2 cutoff appears inconsistent with these ranges; please verify the intended threshold.","section":"Section 4.1"},{"comment":"The KL divergence formula in the supplementary material is typeset as 'yt log yt / ys', which is missing parentheses and the summation over classes; it should be written as sum_i y_t(i) log(y_t(i)/y_s(i)).","section":"Supplementary, Eq. (1)"}],"recommendation":"major_revision","confidential_remarks":"The dataset and benchmark are likely solid contributions, and the paper is generally well organized. The main methodological claim, however, requires additional controlled experiments to rule out confounds from brightness/noise augmentation and distillation. I would encourage the editor to require these experiments before considering the paper for publication, but I do not see the issues as irreparable. The annotation quality information is also needed for a dataset of this scale."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The AODRaw dataset is a real asset. At 7,785 real RAW images, 62 categories, and 9 lighting/weather combinations, it is a substantial step beyond existing RAW detection datasets (RAOD, LOD, PASCALRAW). The annotation effort and the statistical analysis are careful, and the benchmark results across many detectors give the field a useful reference point. The paper deserves credit for demonstrating that a neural ISP adapter is not the only path to good RAW detection performance.\n\nThe soft spot is the central causal claim. The main comparison, RAW pre-training versus sRGB pre-training, changes two variables at once: the input domain and the augmentation schedule. The synthetic ImageNet-RAW is generated by unprocessing with random brightness and noise each iteration, while the sRGB pre-training uses standard ImageNet augmentation. The paper does not include an sRGB pre-training control with the same brightness/noise augmentation. Given that Figures 5 and 6 show sRGB pre-training degrades sharply under brightness/noise shifts, the augmentation alone could plausibly account for part or all of the 1.1% AP gain. The paper also fine-tunes on gamma-corrected real RAW but pre-trains on linear synthetic RAW, which further muddies what \"RAW pre-training\" means.\n\nThis is not a fatal flaw. The dataset is valuable independent of the method claim, and the raw pre-training recipe may well be useful in practice. But the paper's headline result, that the RAW domain itself drives the improvement, is not established without the missing control. Single-run results without error bars add a minor layer of uncertainty. The ablation of distillation is helpful and shows the components do contribute, but it does not isolate the domain effect.\n\nFor a peer-reviewed venue, I would ask for a controlled comparison: sRGB pre-training on ImageNet with the same random brightness/noise augmentation, then fine-tune on AODRaw. Multiple seeds would also help. That experiment is cheap to run and would directly test the stress-test concern. The paper is otherwise well structured and the writing is clear.\n\nRecommendation: send to peer review. It is a dataset plus method paper with real potential value, and the main confound can be resolved with a straightforward experiment. I would bring this to a reading group because the benchmark is useful and the confound is a good teaching point about experimental design.","headline":"Solid new RAW detection benchmark; the pre-training gain is confounded with brightness/noise augmentation, so the causal claim needs a control.","tokens_in":16743,"tokens_out":1485,"would_cite":true,"duration_ms":15063,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Detectors pre-trained on synthetic RAW sensor data outperform sRGB-pretrained models on a new 62-category adverse-weather benchmark, with the largest gains in rain.","keywords":["RAW object detection","AODRaw","adverse conditions","RAW pre-training","cross-domain distillation","unprocessing","object detection benchmark","neural ISP"],"falsifier":"Train the identical detector with sRGB ImageNet pre-training while applying the exact same random brightness and noise augmentation that ImageNet-RAW uses, then fine-tune on AODRaw RAW images; if that control reaches or exceeds 34.8% AP, the claimed RAW-domain advantage is an artifact of the augmentation. A complementary check is to pre-train on a comparably sized set of real RAW images and see whether the gap over sRGB pre-training persists.","tokens_in":15786,"feed_emoji":"🌧️","tokens_out":9566,"duration_ms":77058,"temperature":0.7,"pith_summary":"The paper sets out to establish that object detectors should be trained on unprocessed camera sensor data (RAW) instead of on the 8-bit processed images (sRGB) that conventional datasets use, because the processing step that creates sRGB discards information that matters in rain, fog, and low light. To make this testable, it introduces AODRaw, a dataset of 7,785 real RAW images with 135,601 annotated objects across 62 categories and nine combined light and weather conditions. The central finding is that pre-training on synthetic RAW versions of ImageNet, with knowledge distilled from an sRGB-pretrained teacher, raises Cascade R-CNN with a ConvNeXt-T backbone from 33.7% to 34.8% average precision (AP) on RAW detection, a 1.1-point gain over sRGB pre-training, and the largest gains appear in rain. If this is right, RAW pre-training is a cheap route to more robust object detection that needs no extra image-processing modules.","feed_headline":"Raw-sensor training beats processed images in fog and rain","feed_subtitle":"A 7,785-image RAW benchmark shows pre-training on sensor data improves detection most where weather is worst.","key_machinery":"The load-bearing mechanism is a paired recipe: synthetic ImageNet-RAW pre-training plus cross-domain distillation. Synthetic ImageNet-RAW is produced by the unprocessing method of [2], which reverses an image signal processor to convert sRGB images back to 16-bit RAW-like data and simulates camera noise; because the unprocessing runs inside the data-augmentation pipeline, brightness and noise are randomized each iteration. Cross-domain distillation then trains the RAW-pretrained student with logit-based Kullback-Leibler divergence and feature-based L1 loss against an off-the-shelf sRGB-pretrained teacher of the same architecture, giving the student stable semantic targets that do not vary with the synthesized noise. Together these two pieces let the backbone learn representations that are invariant to brightness and noise before it is fine-tuned on real RAW detection data.","core_discovery":"On its own terms, the paper claims two things. First, the domain gap between sRGB pre-training and RAW fine-tuning is real and costly: a Cascade R-CNN trained on sRGB drops from 34.0% to 28.0% AP when evaluated on RAW, and models pre-trained on sRGB then fine-tuned on RAW underperform models trained and tested entirely in sRGB. Second, pre-training directly in the RAW domain closes most of that gap. Because no large real RAW pre-training set exists, the authors synthesize ImageNet-RAW by unprocessing ImageNet-1K images, inverting the camera pipeline and adding random brightness and shot noise inside the augmentation loop. To help the student cope with noise, they distill logit and feature knowledge from an off-the-shelf sRGB-pretrained teacher. The result is 34.8% AP on AODRaw with Cascade R-CNN and ConvNeXt-T, 1.1 points above sRGB pre-training and 0.8 points above sRGB-based detection, with a 4.8-point gain in rain, achieved without any neural ISP.","pith_inferences":["An implication the paper leaves implicit is that the same unprocessing-plus-distillation recipe should transfer to other sensor modalities that lack large labeled datasets, such as multispectral, high-dynamic-range, or polarization imaging; the sRGB teacher supplies semantics while the synthetic sensor data supplies the input distribution.","The paper does not report an sRGB pre-training control that applies the same random brightness and noise augmentation to ordinary ImageNet-1K; adding that control would decisively separate the benefit of the RAW input distribution from the benefit of the augmentation schedule.","Because the largest gain appears in rain, a natural next experiment is to push the simulated noise and brightness ranges further and test whether detector AP keeps climbing under increasingly extreme synthetic conditions or saturates."],"forward_implications":["A single RAW-pretrained model can serve all nine light and weather conditions at once, eliminating the need for per-condition models or trainable ISP adapters.","The benefit of RAW pre-training is concentrated where it matters most: +0.8 AP in low light, +4.8 AP in rain, and +1.2 AP in fog, compared with +0.9 AP in normal conditions.","The gains are not tied to one architecture: they appear with Cascade R-CNN and ConvNeXt-T in both down-sampled and sliced-image settings, and the raw data supports real-time YOLO-scale detectors without destroying frame rate.","Distillation makes the pre-trained representation measurably more robust to brightness and noise shifts, so downstream fine-tuning starts from a sturdier feature space."],"supporting_citations":[{"why":"Supplies the unprocessing method that converts ImageNet-1K sRGB images into the synthetic 16-bit ImageNet-RAW set used for pre-training.","marker":"[2]"},{"why":"Provides the RAOD baseline and trainable-ISP comparison that RAW pre-training is measured against.","marker":"[34]"},{"why":"Provides the RAW-Adapter neural-ISP baseline that motivates the no-ISP design.","marker":"[8]"},{"why":"Supplies Cascade R-CNN, the detector used in the main comparison and ablation tables.","marker":"[3]"},{"why":"Supplies ConvNeXt-T, the backbone whose pre-training produces the headline 1.1 AP gain.","marker":"[24]"},{"why":"Supplies ImageNet-1K as the source images for synthesizing ImageNet-RAW.","marker":"[29]"},{"why":"Provides the logit-distillation loss used in cross-domain distillation.","marker":"[16]"},{"why":"Provides the feature-distillation loss that combines with logit distillation.","marker":"[32]"},{"why":"Defines the COCO annotation format, evaluation protocol, and category schema that AODRaw follows.","marker":"[21]"}],"fun_headline_variants":["RAW pre-training outperforms sRGB in bad weather detection","Pre-train on RAW, detect better in rain and fog","Synthetic RAW ImageNet boosts detection under adversity","Train on RAW data, not processed images, for robust detection","Direct RAW pre-training closes domain gap in object detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the synthetic ImageNet-RAW images created by unprocessing, with random brightness and simulated shot noise, faithfully represent real camera RAW data well enough that pre-training on them transfers to real AODRaw images; the paper does not include an sRGB pre-training control with the same augmentation schedule, so if that premise fails the reported gains could be an augmentation effect rather than a RAW-domain effect.","fun_headline_variants_meta":{"raw":{"variants":["RAW pre-training outperforms sRGB in bad weather detection","Pre-train on RAW, detect better in rain and fog","Synthetic RAW ImageNet boosts detection under adversity","Train on RAW data, not processed images, for robust detection","Direct RAW pre-training closes domain gap in object detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00088,"raw_usage":{"total_tokens":3819,"prompt_tokens":978,"completion_tokens":2841,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":594,"completion_tokens_details":{"reasoning_tokens":2761}},"tokens_in":594,"tokens_out":2841,"duration_ms":18003,"temperature":1.0,"reasoning_tokens":2761,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:00:58.007452+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the identical detector with sRGB ImageNet pre-training while applying the exact same random brightness and noise augmentation that ImageNet-RAW uses, then fine-tune on AODRaw RAW images; if that control reaches or exceeds 34.8% AP, the claimed RAW-domain advantage is an artifact of the augmentation. A complementary check is to pre-train on a comparably sized set of real RAW images and see whether the gap over sRGB pre-training persists.","supporting_citations":[{"cited_title":"Unprocessing im- ages for learned raw denoising","cited_arxiv_id":null,"evidence_quote":"Supplies the unprocessing method that converts ImageNet-1K sRGB images into the synthetic 16-bit ImageNet-RAW set used for pre-training."},{"cited_title":"Toward raw object detection: A new benchmark and a new model","cited_arxiv_id":null,"evidence_quote":"Provides the RAOD baseline and trainable-ISP comparison that RAW pre-training is measured against."},{"cited_title":"Raw-adapter: Adapting pre- trained visual model to camera raw images","cited_arxiv_id":null,"evidence_quote":"Provides the RAW-Adapter neural-ISP baseline that motivates the no-ISP design."},{"cited_title":"Cascade r-cnn: High quality object detection and instance segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies Cascade R-CNN, the detector used in the main comparison and ablation tables."},{"cited_title":"A convnet for the 2020s","cited_arxiv_id":null,"evidence_quote":"Supplies ConvNeXt-T, the backbone whose pre-training produces the headline 1.1 AP gain."},{"cited_title":"Imagenet large scale visual recognition challenge","cited_arxiv_id":null,"evidence_quote":"Supplies ImageNet-1K as the source images for synthesizing ImageNet-RAW."},{"cited_title":"Microsoft coco: Common objects in context","cited_arxiv_id":null,"evidence_quote":"Defines the COCO annotation format, evaluation protocol, and category schema that AODRaw follows."}],"review_version":1}