{"id":"2332d9a9-f9c0-4245-b655-fc1c59bb2cfe","arxiv_id":"1908.08999","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Photometric normalization before CNN descriptor extraction, especially CLAHE, improves image retrieval across day and night and stays competitive on daylight benchmarks.","lead":"This paper tests whether normalizing image brightness before a retrieval network improves matching across day and night photos, using both classic histogram methods and a learned U-Net. The best method, a hand-crafted contrast-limited histogram equalization, matches or beats the learned network and helps on night queries without hurting standard benchmarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Tokyo 24/7 protocol tests only same-viewpoint illumination change, so the retrieval claim may not generalize to realistic viewpoint variation; the reader's dataset-pair concern is not the load-bearing issue.","rationale":"The strongest claim is empirical: photometric normalization improves day/night retrieval while preserving daylight performance. The evidence for the day/night part rests entirely on the new Tokyo 24/7 protocol. That protocol is a self-defined benchmark that removes viewpoint variation by excluding same-location different-direction images. In a standard retrieval benchmark, positives are not restricted to identical viewpoints; Oxford and Paris contain large viewpoint changes, and the qualitative example in Fig. 1 is a night query against a daylight database with different views. Because the main quantitative day/night result does not involve viewpoint change, the paper does not demonstrate retrieval under the combined illumination+viewpoint variation that occurs in practice. This is not an internal inconsistency; it is an external-validity gap. The reader's weakest_assumption about day/night pair construction is less central: the hand-crafted normalization row 'CLAHE' (Table 1) already improves over VGG GeM without any N/D fine-tuning, so the normalization conclusion is not hostage to the correctness of the reconstructed pairs. The pair validity affects the dataset contribution and the magnitude of the best results, but not the existence of a normalization benefit. I agree with the reader's overall CONDITIONAL verdict, but for a different reason: the condition should be a demonstration on a protocol with viewpoint variation, not just clarification of error bars and clip-limit selection. The paper does include good practices: a control with the same fine-tuning data ('VGG GeM N/D'), a publicly released dataset, and a fair dimensionality comparison in Table 2. These support the internal comparisons. The missing piece is a benchmark that combines illumination and viewpoint change. If the proposed test shows the improvement transfers, the central claim stands; if not, the claim is limited to same-viewpoint matching.","tokens_in":11210,"tokens_out":12732,"duration_ms":130961,"concrete_test":"Run the main comparison (VGG GeM, VGG GeM N/D, CLAHE N/D, U-Net jointly N/D) on the original Tokyo 24/7 place-recognition protocol of Torii et al. [32], i.e., phone night queries against the StreetView database with rendered virtual views, which includes viewpoint variation. If CLAHE N/D and U-Net jointly N/D do not beat VGG GeM N/D by a clear margin on night queries, the claim should be qualified to same-viewpoint illumination change. Alternatively, use the Tokyo 24/7 subset where the first two viewing directions overlap (Fig. 7) and treat those cross-direction same-location pairs as positives; compute mAP for night queries and compare the same methods.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.1 defines a new Tokyo 24/7 protocol in which each query's positives are images from the same location and the same viewing direction under different illumination; same-location different-direction images are explicitly excluded. Thus every positive pair in Table 1 and Table 3 is a near-duplicate viewpoint with only illumination changed (Fig. 7 shows first two directions overlap, second and third do not). This is closer to illumination-invariant verification than to image retrieval, where query and database views typically differ in viewpoint as well. The original Tokyo 24/7 benchmark [32] uses phone-camera queries against StreetView panoramas and rendered virtual views, which includes viewpoint variation; the paper replaces that with a simpler protocol. The central claim in the abstract and Section 5.3(i) is about retrieval under varying illumination, but the quantitative support is generated without viewpoint variation. If the observed gains are specific to same-viewpoint matching, the method may not transfer to standard retrieval. The reader's concern about Retrieval-SfM-N/D pair validity is secondary: the 'CLAHE' row without N/D already outperforms VGG GeM on Tokyo 24/7 (84.1 vs 79.4), so the normalization benefit does not depend on the reconstructed day/night pairs. The load-bearing unproven premise is therefore the evaluation protocol's representativeness.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes photometric normalization as a preprocessing step before CNN descriptor extraction for image retrieval under changing illumination. Three hand-crafted normalizations (gamma correction, histogram equalization, and CLAHE) and a learned U-Net-based normalization are evaluated, with the latter pre-trained on multi-exposure images (See-in-the-Dark) and fine-tuned with a contrastive loss against a VGG-GeM embedding network. A new day/night training dataset, Retrieval-SfM-N/D, is automatically constructed from 3D reconstructions using a ball approximation of co-visible surfaces. Experiments on a newly defined Tokyo 24/7 protocol, plus revisited Oxford and Paris (Medium), show that normalization improves mAP under illumination change, with CLAHE and the learned U-Net performing similarly, while preserving daylight retrieval performance.","tokens_in":11532,"tokens_out":4992,"duration_ms":47184,"significance":"If the results are robust, the paper's contribution is practically valuable: a simple, cheap preprocessing step yields substantial gains for day/night retrieval without hurting standard benchmarks. The experimental design is clear and controlled (same embedding network, three fine-tuning setups, three benchmarks), and the release of the day/night training pairs is a useful resource. The finding that hand-crafted CLAHE matches a learned normalization is also an important and honest observation. However, the central claim is broader than what the new Tokyo 24/7 protocol actually measures, and several methodological details need strengthening before the generality of the claim can be accepted.","major_comments":[{"comment":"The new Tokyo 24/7 protocol excludes same-location different-direction images from the evaluation, so every positive pair is a near-duplicate viewpoint that differs only in illumination. Tables 1 and 3 therefore measure illumination-invariant matching rather than image retrieval under joint viewpoint and illumination change. The original Tokyo 24/7 benchmark [32] includes viewpoint variation through phone-camera queries against StreetView panoramas and rendered virtual views. The abstract and Section 5.3(i) state the claim more generally as \"image retrieval under varying illumination conditions.\" Please either evaluate on the original protocol, add a viewpoint-varying day/night retrieval experiment, or clearly re-scope the claim to same-viewpoint illumination invariance.","section":"Section 5.1"},{"comment":"The construction of Retrieval-SfM-N/D relies on a ball approximation of the visible surface with IoU threshold 0.55 and a 45-degree camera-angle limit, but the paper reports no manual verification or sensitivity analysis of these thresholds. The assertion \"without obvious false positives\" is not supported by quantitative evidence. Because the paper credits the new dataset with the gains described in Section 5.3(ii), the authors should provide validation, such as a sample of inspected pairs or a sensitivity study of the IoU and angle thresholds.","section":"Section 4.1"},{"comment":"The CLAHE clip limit is fixed at 4 and the text states that this \"consistently yielded the best results,\" which suggests the parameter was selected on the evaluation benchmarks. Since CLAHE N/D is one of the top-performing methods in Table 1, the authors should describe a validation protocol (e.g., a hold-out set) or report how the main results vary with the clip limit.","section":"Section 3.1"},{"comment":"All reported results are single runs without error bars or statistical tests. The differences between top methods are small (average mAP 72.4 for CLAHE N/D vs. 72.1 for U-Net jointly N/D vs. 72.9 for Edg+CLAHE N/D in Table 1), yet the discussion in Section 5.3 claims similarities and orderings. Multiple runs or significance statements are needed at least for the headline comparisons that drive the conclusions.","section":"Tables 1-3"}],"minor_comments":[{"comment":"The claim that all image normalization methods outperform the baselines \"by a large margin\" on Tokyo 24/7 is not true for U-Net jointly (79.8) versus VGG GeM (79.4); this statement should be qualified.","section":"Section 5.3(i)"},{"comment":"The caption contains a typo: \"joinlty\" should be \"jointly.\"","section":"Table 1 caption"},{"comment":"The sentence \"the increase was from 4.1GB, 5hrs to 11.6GB, 11hrs\" is unclear; please specify whether these are GPU memory and training time, and for which exact setup.","section":"Section 3.2.1"},{"comment":"The text says \"20 thousand illumination-hard-positive image pairs\" but does not state whether these are unique pairs, how they are split between training and validation, or whether anchors overlap with the Retrieval-SfM positives; please clarify.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The load-bearing issue is the evaluation protocol: the new Tokyo 24/7 protocol is reasonable for studying illumination invariance in isolation, but the paper's general retrieval claim should be re-scoped or supplemented with a viewpoint-varying evaluation. The day/night pair construction is a secondary concern for the normalization claim, since CLAHE without N/D already improves over the baseline, but it matters for the dataset contribution. I would encourage the authors to add either the original Tokyo 24/7 benchmark or a new day/night retrieval set with viewpoint changes, and to report variance estimates for the main tables."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid empirical paper that shows photometric normalization before CNN descriptor extraction helps day/night retrieval while leaving daylight performance roughly intact. The cleanest result is that CLAHE, a simple 8x8 local histogram equalization, lifts Tokyo 24/7 mAP from 79.4 to 84.1 even without fine-tuning on the new day/night pairs, and to 87.0 with them. That is a useful, actionable finding.\n\nWhat is new: the systematic comparison of hand-crafted and learned normalizations for this problem, the Retrieval-SfM-N/D training set, and the per-pair evaluation protocol on Tokyo 24/7. The experiments are well controlled: same VGG-GeM embedding, same fine-tuning recipe, three benchmarks. The paper is honest that the learned U-Net does not beat CLAHE, and that the hand-crafted method is faster.\n\nSoft spots, in order of importance. First, the new Tokyo 24/7 protocol only queries against same-location, same-direction images under different illumination. Same-location different-direction images are excluded. That makes the task closer to illumination-invariant verification than to retrieval, where query and database views usually differ in viewpoint as well. The abstract and Section 5.3(i) make a broader claim about retrieval under varying illumination, but the quantitative support does not include viewpoint variation. The gains could be specific to near-duplicate viewpoints. That is the load-bearing unproven premise, and I think the stress-test note is right on this.\n\nSecond, results are single runs with no error bars. For differences of 1-2 mAP between methods, this matters. Third, the CLAHE clip limit is fixed at 4 because it 'consistently yielded the best results' — effectively selected on the evaluation benchmarks. That is tuning-on-test, albeit mild, and the paper does not report sensitivity.\n\nThe Retrieval-SfM-N/D pair validity concern is real but secondary: the normalization benefit is already visible without the new dataset, so the main retrieval gain does not ride on those reconstructed pairs. Still, a small manual check or sensitivity analysis of the ball-IoU threshold would strengthen the dataset contribution.\n\nBottom line: worth a serious referee. The central empirical direction is sound and the dataset is reusable. I would ask the authors to validate the Tokyo protocol against the original [32] setup with viewpoint-varying queries, add error bars, and justify the CLAHE parameter choice off the test sets.","headline":"Useful preprocessing study with a clean comparison, but the new Tokyo protocol strips out viewpoint variation, so the strong retrieval claim is only partly supported.","tokens_in":12003,"tokens_out":2445,"would_cite":true,"duration_ms":23354,"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":"This paper establishes that photometric normalization before CNN descriptor extraction is enough to make image retrieval work under severe day-to-night illumination changes while preserving daylight retrieval performance.","keywords":["image retrieval","photometric normalization","day-night retrieval","CLAHE","U-Net","illumination invariance","contrastive learning","Retrieval-SfM-N/D"],"falsifier":"Take a random sample of, say, 200 pairs from Retrieval-SfM-N/D and manually check whether the same surfaces are actually visible in both images; if a large fraction are not true same-view pairs, retrain on corrected pairs and see whether the Tokyo 24/7 gains persist. Alternatively, evaluate the same normalisation methods on an independent day/night retrieval benchmark with verified ground truth and compare mAP scores.","tokens_in":11040,"feed_emoji":"🌙","tokens_out":5061,"duration_ms":48107,"temperature":0.7,"pith_summary":"This paper argues that the standard failure of CNN image retrieval on day-to-night queries can be largely fixed by normalizing the photograph before the descriptor is extracted, rather than by changing the retrieval network or its training data alone. Both a hand-crafted normalization, contrast-limited adaptive histogram equalization (CLAHE) applied to the LAB lightness channel, and a learned U-Net normalization improve retrieval on the Tokyo 24/7 day/night benchmark by a large margin over same-dimensionality baselines, while daylight benchmark scores move very little. The learned normalization is pretrained on aligned multi-exposure image pairs and then fine-tuned together with the VGG-GeM embedding using a contrastive loss on a newly constructed dataset of day/night image pairs, Retrieval-SfM-N/D. The paper also shows that the new training pairs alone improve retrieval even without normalization, but the largest gains require the normalization step.","feed_headline":"Normalizing light rescues day-to-night image search","feed_subtitle":"Preprocessing images before CNN descriptors sharply lifts night-query retrieval while daylight search holds.","key_machinery":"The central mechanism is a photometric normalisation stage prepended to a VGG-GeM embedding network. Hand-crafted versions transform the LAB lightness channel by gamma correction, histogram equalisation, or CLAHE; the learned version is a U-Net that takes the RGB image plus a histogram-matched target lightness channel as input and outputs a normalised RGB image. The combined system is trained in a Siamese manner with a contrastive loss on hard-positive pairs, first pretraining the normalisation network on pixel-aligned multi-exposure images and then fine-tuning on the newly constructed day/night pairs of Retrieval-SfM-N/D. The pair construction uses ball approximations of visible 3D points, requiring intersection over union above 0.55 and camera-axis difference below 45 degrees, to find geometrically similar views with large illumination differences.","core_discovery":"The paper's central claim is that a photometric normalisation stage placed before CNN descriptor extraction makes image retrieval robust to severe illumination changes while preserving daytime performance. On a new evaluation protocol built from the Tokyo 24/7 images, every normalisation method examined beats the same-dimensionality VGG GeM and EdgeMAC baselines by a large margin: the best methods, CLAHE N/D and U-Net jointly N/D, reach Tokyo mAP scores of 87.0 and 86.5 against a 79.4 VGG GeM baseline, while Revisited Oxford and Paris Medium scores stay within about one point of the baseline. The learned U-Net normalisation matches but does not beat simple CLAHE, which the paper reads as evidence that once the input is properly normalised, the embedding network can extract the same discriminative information. Fine-tuning on the new day/night pairs helps both with and without normalisation, but the normalisation stage is what delivers the largest illumination robustness gains.","pith_inferences":["Since the learned U-Net matches but does not beat CLAHE, the results suggest the bottleneck for illumination-robust retrieval is not the form of the normalisation but the existence of a normalisation stage at all; cheap hand-crafted normalisation may be sufficient in practice.","A natural extension the authors do not run is pretraining the normalisation on other aligned multi-exposure data and fine-tuning on day/night pairs from other 3D reconstructions, which would test whether the gains transfer to city-scale or indoor scenes.","Because the method operates at the preprocessing level, it could be plugged into any CNN descriptor or downstream task such as place recognition or low-light person re-identification, not only the VGG-GeM retrieval pipeline tested here."],"forward_implications":["On the proposed Tokyo 24/7 protocol, adding CLAHE or learned U-Net normalisation raises mAP from 79.4 (VGG GeM) to 87.0 or 86.5, a large gain at the same descriptor dimensionality.","Fine-tuning on the new Retrieval-SfM-N/D day/night pairs improves retrieval even without normalisation, but normalisation remains necessary for the largest gains.","Daylight retrieval performance is largely preserved: ROxf/RPar Medium scores for CLAHE N/D (60.2/70.0) and U-Net jointly N/D (60.2/69.6) sit close to the VGG GeM baseline (60.9/69.3).","Combining normalisation with an EdgeMAC+VGG ensemble improves results further, and day-night query pairs are the ones that benefit most from the method."],"supporting_citations":[{"why":"Supplies the See-in-the-dark aligned multi-exposure image pairs used to pre-train the U-Net normalisation network.","marker":"[3]"},{"why":"Provides the VGG-GeM embedding network, the fine-tuning procedure with contrastive loss, and the Retrieval-SfM clusters used for training and whitening.","marker":"[25]"},{"why":"EdgeMAC baseline, a partially illumination-invariant edge descriptor that is compared and also combined with VGG-GeM descriptors.","marker":"[24]"},{"why":"Defines the revisited Oxford and Paris datasets and evaluation protocol used to check that daylight retrieval performance is preserved.","marker":"[22]"},{"why":"Introduces the Tokyo 24/7 day, sunset and night phone-camera images that the new illumination-retrieval protocol is built on.","marker":"[32]"},{"why":"U-Net architecture that the learned normalisation network is based on.","marker":"[26]"},{"why":"Provides the U-Net implementation that the authors adapt by replacing the final tanh layer and removing batch normalisation.","marker":"[6]"},{"why":"Very deep convolutional network used to initialise the VGG embedding network before retrieval fine-tuning.","marker":"[27]"}],"fun_headline_variants":["Pre-CNN light fix boosts day-night image retrieval","Simple normalization sharpens night image search","Normalize first, then CNN: day-night retrieval gains","CLAHE equals learned U-Net for illumination-robust retrieval"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole day/night training signal rests on the newly built Retrieval-SfM-N/D pairs, which are accepted when the overlap of ball approximations of visible 3D points exceeds 0.55 and camera axes differ by at most 45 degrees; if many of these pairs do not actually show the same scene surface, the reported gains from N/D fine-tuning could come from something else.","fun_headline_variants_meta":{"raw":{"variants":["Pre-CNN light fix boosts day-night image retrieval","Simple normalization sharpens night image search","Normalize first, then CNN: day-night retrieval gains","CLAHE equals learned U-Net for illumination-robust retrieval"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1343,"prompt_tokens":863,"completion_tokens":480,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":479,"completion_tokens_details":{"reasoning_tokens":416}},"tokens_in":479,"tokens_out":480,"duration_ms":6375,"temperature":1.0,"reasoning_tokens":416,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:23:54.192220+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random sample of, say, 200 pairs from Retrieval-SfM-N/D and manually check whether the same surfaces are actually visible in both images; if a large fraction are not true same-view pairs, retrain on corrected pairs and see whether the Tokyo 24/7 gains persist. Alternatively, evaluate the same normalisation methods on an independent day/night retrieval benchmark with verified ground truth and compare mAP scores.","supporting_citations":[{"cited_title":"Learning to see in the dark","cited_arxiv_id":null,"evidence_quote":"Supplies the See-in-the-dark aligned multi-exposure image pairs used to pre-train the U-Net normalisation network."},{"cited_title":"Fine- tuning CNN image retrieval with no human annotation","cited_arxiv_id":null,"evidence_quote":"Provides the VGG-GeM embedding network, the fine-tuning procedure with contrastive loss, and the Retrieval-SfM clusters used for training and whitening."},{"cited_title":"Deep shape matching","cited_arxiv_id":null,"evidence_quote":"EdgeMAC baseline, a partially illumination-invariant edge descriptor that is compared and also combined with VGG-GeM descriptors."},{"cited_title":"Revisiting Oxford and Paris: Large-scale image retrieval benchmarking","cited_arxiv_id":null,"evidence_quote":"Defines the revisited Oxford and Paris datasets and evaluation protocol used to check that daylight retrieval performance is preserved."},{"cited_title":"24/7 place recognition by view synthesis","cited_arxiv_id":null,"evidence_quote":"Introduces the Tokyo 24/7 day, sunset and night phone-camera images that the new illumination-retrieval protocol is built on."},{"cited_title":"Image-to-image translation with conditional adver- sarial networks","cited_arxiv_id":null,"evidence_quote":"Provides the U-Net implementation that the authors adapt by replacing the final tanh layer and removing batch normalisation."},{"cited_title":"Very deep convo- lutional networks for large-scale image recognition","cited_arxiv_id":null,"evidence_quote":"Very deep convolutional network used to initialise the VGG embedding network before retrieval fine-tuning."}],"review_version":1}