{"id":"e012d15f-87ac-4c7e-ab7c-990a1ffac78c","arxiv_id":"2412.07612","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ViewDelta uses text prompts to define relevant scene changes, enabling one model to work across multiple change-detection datasets and view angles.","lead":"A new system, ViewDelta, detects scene changes in image pairs based on a text prompt that tells it which changes matter, and can be trained on many different datasets at once. The authors also release a half-million-pair synthetic dataset, CSeg, to support text-conditioned change detection.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unaligned variants (Diff-1/Diff-2) are trained and evaluated on without stated split provenance; if test images leaked into training, Tables 4 and 5 inflate ViewDelta's viewpoint robustness and the central generalization claim.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the unaligned variants are used both for training and evaluation without explicit split provenance. This is the most serious issue because it directly undermines the headline empirical claim of generalizable, view-robust SCD. If the variants were derived from the full datasets, the large margins over Dinov2 RSCD in Tables 4 and 5 are explained by test set leakage rather than by ViewDelta's design. The paper provides no statement that the original train/test splits were preserved when the perturbed variants were created, and the reference to Lin et al. [28] does not clarify this in the text. Other concerns (no error bars on main tables, hand-crafted prompts, limited manual validation of CSeg) are secondary: they affect interpretability or robustness but do not, if true, invalidate the central comparison. The data-provenance issue, if true, invalidates the central comparison. Since the concern is not confirmed—it could be that the variants were correctly split—the appropriate verdict remains CONDITIONAL, matching the reader's judgment. I propose no change to the verdict, but the authors must resolve the provenance question and, if needed, correct Tables 4 and 5.","tokens_in":14017,"tokens_out":2684,"duration_ms":22846,"concrete_test":"Inspect the released code/data generation scripts from Lin et al. [28] (arXiv:2409.16850) to verify that the Diff-1 and Diff-2 variants of PSCD and VL-CMU-CD were created only from the original training splits, with the original test pairs excluded. Concretely, for every image pair in the ViewDelta test sets of Tables 4 and 5, confirm that neither the image itself nor any perturbed (Diff-1/Diff-2) version appears in the ViewDelta training list. Alternatively, retrain ViewDelta with the variants split by the original train/test partition and re-evaluate; if the F1 scores on PSCD Diff-2 or VL-CMU-CD Diff-2 drop by more than 5 points, the reported viewpoint robustness is contaminated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a single ViewDelta model trained jointly on CSeg, SYSU-CD, PSCD, VL-CMU-CD, and the unaligned variants of PSCD and VL-CMU-CD performs competitively or better than dataset-specific models, including under viewpoint changes. Section 3.6 states that training uses the Diff-1 and Diff-2 variants, and Section 5.1 evaluates on exactly those same variants (Tables 4 and 5). The paper never states that the variants were constructed only from the original training splits of PSCD and VL-CMU-CD. If Lin et al. [28] generated the perturbed image pairs from the full datasets—including test images—then test-time images or their perturbed versions were seen during training. Because the perturbed pairs are derived from the originals via nearest-neighbor matching and parallax/occlusion effects, they are near-duplicates of the test images, making the reported F1 scores (e.g., 63.6 vs. Dinov2 RSCD's 28.4 on PSCD Diff-2) artificially high. The claim of 'robustness to viewpoint changes' is therefore load-bearing on an unverified data-provenance assumption. This is not a matter of disagreement with consensus; it is an internal inconsistency in the evaluation protocol if the assumption fails. The paper's own text does not resolve the ambiguity, and no supplementary statement or code reference clarifies the split derivation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents ViewDelta, a text-conditioned scene change detection (SCD) framework that takes two images and a natural-language prompt and outputs a binary change mask. To enable this, the authors release CSeg, a large synthetic dataset with over 500K image pairs and 300K unique prompts. ViewDelta is trained jointly on CSeg, SYSU-CD, PSCD, VL-CMU-CD, and the unaligned Diff-1/Diff-2 variants of PSCD and VL-CMU-CD. Experiments report competitive or superior performance against dataset-specific baselines, particularly on viewpoint-challenging variants, plus ablations isolating the contribution of frozen image/text embeddings and segmentation query tokens.","tokens_in":14147,"tokens_out":9033,"duration_ms":72577,"significance":"If the reported results are reliable, ViewDelta demonstrates a promising direction: using text prompts to disambiguate change relevance across datasets with conflicting labeling conventions, enabling a single model to work across domains and view changes. The CSeg dataset is a substantial resource. The paper includes useful ablations and a reproducible architecture. However, the headline viewpoint-robustness numbers (Tables 4–5) rest on an unverified data-provenance assumption about the unaligned variants, and the novelty claim is weakened by unaddressed prior work on text-conditioned change detection. These issues need to be resolved before the contribution can be fully assessed.","major_comments":[{"comment":"The paper trains ViewDelta on the Diff-1 and Diff-2 unaligned variants of PSCD and VL-CMU-CD and evaluates on the same variants, but it never specifies whether those variants were created solely from the original training splits. If Lin et al. [28] generated the variants from the full datasets including test images, then test-time images or near-duplicates thereof would have been seen during training, inflating the reported F1 scores (e.g., 63.6 vs. 28.4 for Dinov2 RSCD on PSCD Diff-2) and invalidating the claim of robustness to viewpoint changes. The manuscript must state and substantiate the exact split provenance, or re-run the evaluation using variants derived only from training splits.","section":"Section 3.6 and Section 5.1, Tables 4 and 5"},{"comment":"The paper claims as a contribution 'a novel text prompt conditioned change detection task,' but prior work already exists in this area—notably Sachdeva and Zisserman's 'The Change You Want to See' [39] and its 3D extension [40], which introduce language-conditioned change detection together with a synthetic dataset (COCO-Inpainted). The authors cite these works only as synthetic dataset generation and do not compare against or discuss them as text-conditioned methods. This weakens the novelty claim; the authors should position ViewDelta relative to these methods and include empirical comparisons on a common benchmark (e.g., CSeg or COCO-Inpainted).","section":"Section 1 (Introduction) and Section 2 (Related Work)"},{"comment":"ViewDelta results are reported as point estimates without error bars or repeated-run variance, while the Gemini baseline in Table 1 includes 95% confidence intervals. On SYSU-CD (Table 3), the differences between ViewDelta and state-of-the-art baselines are within a few IoU points (e.g., 70.09 vs. 71.10 for MambaBCD; 67.05 vs. 66.79 for TransUNetCD), so without variance estimates it is unclear whether these differences are meaningful. The authors should report confidence intervals or multiple seeds, at least for the main comparisons.","section":"Section 5.1, Tables 2–5"},{"comment":"The evaluation prompts for SYSU-CD and VL-CMU-CD are long, hand-crafted lists of classes or descriptions (e.g., 'urban development, suburban expansion, pre-construction groundwork, vegetation alteration, road widening, and coastal construction'). The manuscript provides no sensitivity analysis or protocol for selecting these prompts, leaving open the possibility that they were chosen to maximize performance. The authors should either use a validation set to select prompts or report results across multiple reasonable prompt variants; otherwise the fairness of these comparisons cannot be assessed.","section":"Section 5.1, Tables 3 and 5"}],"minor_comments":[{"comment":"The CSeg statistics are inconsistent: the test set contains 7,326 unique prompts, yet the text reports 35,271 unique prompts not seen in train. Please clarify the counts.","section":"Section 4.3"},{"comment":"The dataset URL appears as 'github.io/viewdelta/' in the abstract and as 'https://joshuakgao.github.io/viewdelta/' in the contributions; the correct URL should be used consistently.","section":"Abstract"},{"comment":"ViewDelta is evaluated on the full CSeg test set, while Gemini 2.5 Pro is evaluated on a 2000-sample subset; the sampling procedure should be described and the uneven comparison noted in the table caption.","section":"Table 1"},{"comment":"ViewDelta's fine-tuned F1 of 63.1 on aligned PSCD is far above the previous best of 44.2 (Dinov2 RSCD); the paper does not analyze this large gap, and a discussion of possible reasons would strengthen the work.","section":"Table 4"},{"comment":"The sentence 'We investigate directly using image features from a frozen Dinov2 model as embeddings' is ambiguous; please specify that these features replace the learned patch embeddings.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The data-provenance concern about the Diff-1/Diff-2 variants is, in my view, the decisive issue for the viewpoint-robustness claim. The authors may have a straightforward explanation (e.g., Lin et al. only used training splits), but the paper must state it explicitly. The novelty question regarding text-conditioned change detection should also be checked carefully during revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know about this one. First, it is not only an architecture paper: CSeg, the 500k-pair synthetic text-conditioned change detection dataset, is a real resource, and the joint multi-dataset training protocol is a sensible way to reconcile conflicting label conventions. Second, the paper's most striking numbers—the large jumps on the unaligned PSCD and VL-CMU-CD variants—rest on a data-provenance assumption the paper never states. We are told the Diff-1/Diff-2 variants from Lin et al. are used for training and evaluation, but never that those variants were generated only from the original training splits. If they were generated from the full datasets, test images (or near-duplicates through perturbation) were in the training set, and Tables 4 and 5 would be inflated. This is not a nitpick; it is the load-bearing part of the generalization claim. The stress-test note convinced me this is a legitimate ambiguity, and the authors need to resolve it.\n\nWhat is actually good: the CSeg generation pipeline is thoughtful—red herring masks, prompt templates, 'all' prompts, and reported train/test prompt leakage statistics. The ablations show frozen Dinov2 features and segmentation query tokens matter, and the w/o-prompt ablation shows text conditioning is doing real work, not architectural padding. The empirical claim that a single model can be competitive with dataset-specific baselines holds for SYSU-CD and VL-CMU-CD, and even on PSCD the general model is only a few points behind fine-tuned.\n\nSoft spots beyond the provenance issue: no direct comparison with prior text-conditioned SCD methods (COCO-Inpainted / 'The Change You Want to See'), which is awkward because that work is cited as inspiration. No error bars on ViewDelta's own numbers. Evaluation prompts are hand-tuned per dataset. CSeg's label quality check is 500 images, 0.1% of the data. None of these individually is fatal; the provenance question is the one that matters.\n\nBottom line: this deserves a serious referee. If the split provenance checks out, it is a solid contribution with a useful dataset. I'd cite the CSeg dataset if I worked in this area. Bring it to reading group.","headline":"Strong dataset and sensible framework; the headline generalization numbers hinge on a split-provenance detail the paper never documents.","tokens_in":14868,"tokens_out":3586,"would_cite":true,"duration_ms":29273,"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":"ViewDelta claims that a single text-conditioned model, trained jointly on several scene change detection datasets, performs as well as or better than models trained per dataset, and that text conditioning resolves the labeling conflicts…","keywords":["scene change detection","text-conditioned segmentation","vision-language models","joint training","synthetic dataset","viewpoint invariance","CSeg dataset","generalization"],"falsifier":"Inspect the Diff-1 and Diff-2 test splits of PSCD and VL-CMU-CD for any image whose original or perturbed version appears in the training splits; if an overlap exists, the F1 scores in Tables 4 and 5 are inflated by train/test leakage and the viewpoint-robustness claim weakens.","tokens_in":13627,"feed_emoji":"🔍","tokens_out":6895,"duration_ms":59372,"temperature":0.7,"pith_summary":"ViewDelta is a framework for scene change detection that lets a user define, in natural language, which changes count as relevant before the model looks at an image pair. The paper argues that this text conditioning dissolves the labeling conflicts that have kept scene change detection models from training jointly on multiple datasets, because a prompt such as \"vehicle\" or \"all observable differences\" tells the model which of the dataset's labels apply. To support the argument, the authors build CSeg, a synthetic dataset of over 500,000 image pairs with more than 300,000 unique prompts, and train a single ViewDelta model jointly on CSeg, PSCD, SYSU-CD, VL-CMU-CD, and unaligned variants of the latter two. They report that this one model is competitive with or better than dataset-specific baselines, including on unaligned image pairs with viewpoint change. If the results hold, text conditioning is a practical route to a single generalizable change detector rather than a separate model per benchmark.","feed_headline":"A single text-prompted model matches specialized change detectors","feed_subtitle":"ViewDelta trains once on five change-detection datasets; a sentence at inference decides what counts as change.","key_machinery":"The machinery is a prompt-conditioned transformer: a frozen SigLip text encoder turns the prompt into tokens, a frozen Dinov2 vision transformer embeds the before and after images, and these tokens are concatenated with learnable segmentation query tokens and passed through a ViT backbone. The segmentation head reads only the segmentation query tokens, deliberately avoiding operations such as difference or concatenation of pixel-aligned features that assume the two views are spatially aligned. The other load-bearing piece is the CSeg generation procedure, which uses large vision-language model class proposals and open-vocabulary segmentation masks, inpainting to create changes, affine transforms for viewpoint variation, and \"red herring\" masks so the model learns to ignore inpainting artifacts and follow the text.","core_discovery":"The paper's central claim is that scene change detection can be made generalizable by conditioning the binary change mask on a text prompt that specifies the relevant change, and that this conditioning is strong enough to permit one ViewDelta model to be trained jointly across datasets whose labeling conventions contradict each other. On CSeg, the general model reaches 83.80 IoU; on PSCD multi-class scene change detection it reaches 51.2 IoU without dataset-specific training, within 4.3 IoU of its own fine-tuned version; on SYSU-CD it reaches 67.05 IoU; and on the unaligned Diff-1/Diff-2 variants of PSCD and VL-CMU-CD it stays close to its aligned performance. The authors interpret these numbers as showing that text prompts resolve the ambiguity of what counts as change, that avoiding spatial alignment assumptions in the segmentation head is what preserves performance under viewpoint changes, and that the general model is a good starting point for further fine-tuning.","pith_inferences":["An untested implication is that ViewDelta inherits an open vocabulary from its frozen text encoder, so zero-shot evaluation on prompts like \"damage\" or \"flooding\" over unlabeled image pairs would reveal whether the generalization claim extends beyond the benchmark class lists.","The CSeg recipe of vision-language class proposals, segmentation masks, inpainting, and red herring masks could be reused to produce text-conditioned change data in domains such as medical imaging or aerial surveying, where no such dataset currently exists.","Because the segmentation head never aligns the two image streams, the same architecture may extend to video frames with large camera motion, a setting where most change detectors fail; this is an inference, not something the paper evaluates.","The large margin over the Gemini baseline on CSeg may partly reflect that the synthetic prompts are near the training distribution; a harder test would be human-authored prompts on real image pairs with viewpoint differences."],"forward_implications":["A single ViewDelta model can serve street-view, satellite, and multi-view change-detection tasks at inference time, with the user's text deciding what counts as a change.","Future scene change detection datasets can be pooled without manually reconciling their label definitions, because the prompt supplies the missing context for each training example.","The model keeps most of its accuracy when the input pair is unaligned, indicating that real-world change detection from casually captured, non-registered photos is within reach.","Fine-tuning the jointly trained model gives a consistent boost over training per dataset, so text-conditioned joint pretraining is a reusable starting point for specialized change detectors."],"supporting_citations":[{"why":"Supplies the Diff-1 and Diff-2 unaligned variants of PSCD and VL-CMU-CD used for training and evaluation, plus the Dinov2 RSCD baseline that ViewDelta must beat.","marker":"[28]"},{"why":"Provides the PSCD street-view dataset with eight semantic classes used in joint training and multi-class evaluation.","marker":"[42]"},{"why":"Provides the SYSU-CD satellite change detection dataset used for joint training and cross-domain evaluation.","marker":"[43]"},{"why":"Provides the VL-CMU-CD street-view dataset with label noise used for joint training and unaligned evaluation.","marker":"[1]"},{"why":"Supplies the SA-1B images and masks that are the raw material for the CSeg generation procedure.","marker":"[26]"},{"why":"Establishes the inpainting-plus-affine-transformation method for synthetic change pairs and the red herring strategy that CSeg adapts.","marker":"[40]"},{"why":"Produces the object class proposals that become the diverse text prompts in CSeg.","marker":"[29]"},{"why":"Provides the LaMa inpainting model used to simulate the after-image changes in CSeg.","marker":"[48]"},{"why":"Provides the frozen SigLip text encoder whose generalization capacity the paper relies on for open-vocabulary prompts.","marker":"[54]"},{"why":"Provides the frozen Dinov2 image embeddings that the ablation shows are crucial for data-efficient joint training.","marker":"[32]"}],"fun_headline_variants":["One model, any change: text prompts define what counts as change","Say what changed: single model matches specialized change detectors","Text prompts make change detection generalizable across datasets","ViewDelta: words define what changes, one model for all","Language conditions change detection: one model, many datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results on unaligned benchmarks rest on the unstated assumption that the Diff-1 and Diff-2 variants of PSCD and VL-CMU-CD used for training were built only from the original training splits, so no test image or a perturbed version of one was seen during training.","fun_headline_variants_meta":{"raw":{"variants":["One model, any change: text prompts define what counts as change","Say what changed: single model matches specialized change detectors","Text prompts make change detection generalizable across datasets","ViewDelta: words define what changes, one model for all","Language conditions change detection: one model, many datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00108,"raw_usage":{"total_tokens":4529,"prompt_tokens":965,"completion_tokens":3564,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":3494}},"tokens_in":581,"tokens_out":3564,"duration_ms":26988,"temperature":1.0,"reasoning_tokens":3494,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:39:46.044554+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the Diff-1 and Diff-2 test splits of PSCD and VL-CMU-CD for any image whose original or perturbed version appears in the training splits; if an overlap exists, the F1 scores in Tables 4 and 5 are inflated by train/test leakage and the viewpoint-robustness claim weakens.","supporting_citations":[{"cited_title":"Weakly supervised silhouette-based semantic scene change detec- tion, 2022","cited_arxiv_id":null,"evidence_quote":"Provides the PSCD street-view dataset with eight semantic classes used in joint training and multi-class evaluation."},{"cited_title":"A deeply supervised attention metric-based network and an open aerial image dataset for remote sensing change detection","cited_arxiv_id":null,"evidence_quote":"Provides the SYSU-CD satellite change detection dataset used for joint training and cross-domain evaluation."},{"cited_title":"Street-view change detection with deconvolutional networks","cited_arxiv_id":null,"evidence_quote":"Provides the VL-CMU-CD street-view dataset with label noise used for joint training and unaligned evaluation."},{"cited_title":"Segment any- thing","cited_arxiv_id":null,"evidence_quote":"Supplies the SA-1B images and masks that are the raw material for the CSeg generation procedure."},{"cited_title":"The change you want to see (now in 3d)","cited_arxiv_id":null,"evidence_quote":"Establishes the inpainting-plus-affine-transformation method for synthetic change pairs and the red herring strategy that CSeg adapts."},{"cited_title":"Dinov2: Learning robust visual features with- out supervision, 2024","cited_arxiv_id":null,"evidence_quote":"Provides the frozen Dinov2 image embeddings that the ablation shows are crucial for data-efficient joint training."}],"review_version":1}