{"id":"c31f7037-2b96-48d8-a77b-ea6dfb111ecd","arxiv_id":"2505.04758","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"SATNet combines a decoupled attention module and dual texture/saliency pyramids for RGB-D salient object detection, reporting strong benchmark accuracy while omitting the depth-estimation model from its efficiency accounting.","lead":"This paper proposes a lightweight RGB-D salient object detection network that replaces sensor depth maps with pseudo depth maps generated by the Depth Anything Model. The reported 5.2M parameters and 415 FPS exclude this external depth estimator, which undercuts the paper's speed-accuracy tradeoff claim.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Efficiency claim omits the Depth Anything estimator that is required at inference and shown inside the framework, so the advertised 5.2M-parameter, 415-FPS tradeoff is not measured for the full system.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: Depth Anything is silently excluded from the efficiency accounting and from the comparison with raw-depth baselines. My read does not change the reader's verdict. The concern is not that SATNet's modules are internally unsound; the ablations in Tables VI-X support the contribution of DAM, DIRM, and DFAM, and Table V shows SATNet still outperforms two lightweight baselines even when all are trained on pseudo depth. But the central claim in the abstract is explicitly an efficiency claim, and that claim is incomplete because the depth estimator is both required at inference and depicted as part of the framework in Fig. 2. Omitting its parameters, FLOPs, and inference time from Table III and from the headline 5.2M/415FPS makes the speed-accuracy tradeoff unverifiable as stated. The accuracy improvements may remain valid under the paper's protocol, but the headline does not hold for the full system. A revised version that reports end-to-end efficiency with a stated Depth Anything variant and reruns comparisons under matched depth inputs could support a conditional acceptance, but as presented the reader's REJECT stands.","tokens_in":25209,"tokens_out":3783,"duration_ms":42197,"concrete_test":"Run the released code with a timing script that executes Depth Anything followed by SATNet on the same GPU and input resolution used for the 415 FPS figure, and report end-to-end FPS and total parameter count including the Depth Anything checkpoint; if end-to-end FPS is materially below 415 FPS or parameter count is above 5.2M, the advertised lightweight efficiency claim fails as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central contribution is a 'lightweight framework with 5.2 M parameters and 415 FPS.' However, the framework in Fig. 2 and in Section III-A sends RGB images through the Depth Anything Model to generate pseudo depth, and the final model is trained and evaluated on pseudo depth (Table IV, case d). Depth Anything is therefore part of the deployed inference pipeline, not optional preprocessing: without it, no depth input exists for the depth encoder. Table III reports per-stage time and parameters for Encoder, DAM, DIRM, and Decoder only, and the abstract/header efficiency numbers omit the depth estimator. Even the smallest Depth Anything variants add tens of millions of parameters and substantial per-image computation, so the stated 5.2M and 415 FPS describe only the downstream network. The accuracy comparison is similarly not normalized: SATNet uses pseudo depth while most baselines in Table I use low-quality dataset depth, although Table V partially controls for this by retraining AirSOD and MobileSal on pseudo depth. This is not an attack on the module design; the ablations and released code are real supporting evidence. The load-bearing issue is that the speed-accuracy tradeoff, as advertised and titled, cannot be reproduced or fairly compared until the depth estimator's cost is included and the depth input source is controlled.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SATNet, a lightweight RGB-D salient object detection network built on two MobileNet-v2 encoders, a Decoupled Attention Module (DAM), a Dual Information Representation Module (DIRM) with texture and saliency branches, and a decoder containing Dual Feature Aggregation Modules (DFAM). The Depth Anything Model is used to replace low-quality dataset depth maps with pseudo depth maps for both training and testing. The authors report strong accuracy on five RGB-D SOD benchmarks and claim a lightweight framework with 5.2 M parameters and 415 FPS, along with extensive ablations, failure cases, and extensions to polyp segmentation and RGB-T SOD.","tokens_in":1366,"tokens_out":3118,"duration_ms":94474,"significance":"If the reported efficiency numbers described the actual deployed system, the paper would offer a genuinely useful speed-accuracy tradeoff analysis: the module designs are concrete, the ablations are extensive, and the released code is a real strength. The DAM comparison against standard attention mechanisms and the DFAM comparison against ASPP/DenseASPP are informative, and Table V provides a partial control for the pseudo-depth input. However, the central efficiency claim is not measured for the full inference pipeline because the Depth Anything estimator is excluded from parameter, FLOPs, and FPS counts, and the accuracy comparison against raw-depth baselines is not controlled for the depth input source. These issues affect the headline contribution and require substantial revision.","major_comments":[{"comment":"The headline claims of 5.2 M parameters and 415 FPS describe only the downstream SATNet, not the full system. In Fig. 2 and III-A, the Depth Anything Model is applied to every RGB input to produce the pseudo depth map consumed by the depth encoder, and the final model is trained and tested on pseudo depth (Table IV, case d). Without this model, the depth encoder has no input at test time. Table III lists per-stage time and parameters only for Encoder, DAM, DIRM, and Decoder, and Table I does not include the Depth Anything Model in Params, FLOPs, or FPS. The smallest Depth Anything variants add tens of millions of parameters and substantial per-image latency, so the stated 5.2 M/415 FPS cannot be reproduced for the system whose speed-accuracy tradeoff is the paper's title and central contribution. Please report full-system cost including the depth estimator, or explicitly reposition the claims as applying only to the downstream network and state the full-system numbers in the abstract and Table I.","section":"Abstract; III-A; Tables I and III"},{"comment":"The accuracy comparison against SOTA baselines is not on equal footing. SATNet is trained and evaluated with pseudo depth from Depth Anything, while the baselines in Table I are evaluated with the low-quality dataset depth (or their own depth inputs). Table V shows that retraining MobileSal and AirSOD on pseudo depth improves their SIP MAE from 0.057 to 0.052 and from 0.060 to 0.054, respectively, so a nontrivial part of SATNet's advantage over these lightweight methods may come from the input representation rather than the proposed architecture. The claim in IV-D that SATNet 'excels SOTA CNN-based heavyweight models' is not supportable until the baselines are given the same pseudo-depth input or the comparison is otherwise controlled. At minimum, all lightweight baselines in Table I should be retrained with pseudo depth, and the speed comparison should include the cost of generating pseudo depth.","section":"IV-D, Table I; IV-E1, Table V"},{"comment":"The hyperparameters psi_ef and the DFAM kernel/dilation combinations are selected by evaluating on the SIP and NLPR test sets, which are the same datasets used for the final headline numbers in Tables I, VI, VII, and VIII. The paper does not describe a validation split or cross-validation for these choices. This is test-set tuning: values such as psi_ef=32 and CK&DR=(3,5,7) are chosen to optimize the reported metrics, so the final results are not an independent evaluation of the configuration. Please introduce a held-out validation protocol for hyperparameter selection and report final results on a genuine test protocol, or clearly acknowledge this tuning and temper the corresponding claims.","section":"IV-E4 and IV-E5; Tables IX and X"}],"minor_comments":[{"comment":"The sentence 'Ours vs AirSOD: 365 vs 415' reverses the order; Table I lists AirSOD at 365 FPS and SATNet at 415 FPS.","section":"IV-D1"},{"comment":"The row 'Ours-Conv3' contains an extra 0.045 and appears to omit the NLPR S-measure value; please correct the table.","section":"Table VI"},{"comment":"The sentence 'we focus on lightweight RGB SOD models' should presumably read 'lightweight RGB-D SOD models'.","section":"II-A"},{"comment":"The module is introduced as Global Semantic Refinement (GSR), but the ablation text and Table VII also use 'GRS'; please standardize the abbreviation.","section":"III-C and Table VII"},{"comment":"The phrase 'to optimal its parameters' should be 'to optimize its parameters'.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The module-level contributions are real and the code release is a positive factor. The decisive question for the editor is whether the authors can provide corrected, full-system efficiency numbers (including Depth Anything) and controlled accuracy comparisons without losing the paper's central speed-accuracy tradeoff claim. If the full-system FPS and parameter count are dramatically worse, the framing of the paper may need to change. The test-set tuning issue also needs to be addressed in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper's headline number isn't measuring what it claims. The 5.2M/415FPS figure covers only the downstream network, while the Depth Anything model that generates the pseudo depth maps is drawn inside the framework and is required to produce the depth input. That makes the speed-accuracy tradeoff claim, as advertised, unsupported.\n\nWhat the paper does well: it is a competent engineering effort. The three modules (DAM, DIRM, DFAM) are believable, and the ablation study is unusually thorough: depth quality is tested with train/test combinations (Table IV), and the authors re-train MobileSal and AirSOD on pseudo depth to partly control the comparison (Table V). That is good practice, and the code is released. The empirical pattern—pseudo depth from Depth Anything helping these SOD models—is plausible and useful on its own.\n\nWhere it falls down: the efficiency accounting. The pipeline in Fig. 2 includes the Depth Anything Model, and Table IV case (d) shows the best results come from using pseudo depth in both training and testing. So at inference, either you run Depth Anything (large, slow) or you have precomputed pseudo depth offline. In either case, the comparison against raw-depth baselines in Table I is not a like-for-like tradeoff. The authors hide this by reporting per-stage times for only Encoder/DAM/DIRM/Decoder (Table III). The mismatch is not a minor omission; it changes the headline conclusion. Even a small Depth Anything variant adds tens of millions of parameters and substantial latency.\n\nA weaker but real soft spot: some hyperparameters (kernel/dilation rates, efficiency factor) were tuned on the SIP and NLPR test sets used for final numbers. That tends to inflate results slightly. It is common in the field, but worth noting.\n\nBottom line: the accuracy results under the paper's protocol are probably fine, and the module design looks sound. The problem is the framing of the efficiency claim. I would not desk-reject it, but a serious referee should ask for a version that either includes the depth estimator in the parameter/FPS counts or explicitly presents the method as 'pseudo-depth precomputed' and then compares only against pseudo-depth baselines. As written, the 'lightweight 5.2M/415FPS' claim does not hold for the full system. The paper is worth a reading-group discussion on fair efficiency reporting.","headline":"The advertised 5.2M/415FPS describes only the downstream network, not the Depth Anything model that generates the pseudo depth, so the central speed-accuracy claim is not measured for the full system.","tokens_in":25976,"tokens_out":4253,"would_cite":false,"duration_ms":39770,"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":"A 5.2-million-parameter RGB-D salient object detector, SATNet, beats heavyweight CNN models on five benchmarks while running at 415 frames per second.","keywords":["RGB-D salient object detection","lightweight network","pseudo depth","Depth Anything Model","decoupled attention","dual information representation","speed-accuracy tradeoff","real-time inference"],"falsifier":"Measure the full test-time pipeline, including the Depth Anything forward pass on each RGB image, on the same GPU used to report 415 FPS; if the end-to-end throughput is materially lower, the headline speed does not describe the deployable system.","tokens_in":24988,"feed_emoji":"⚡","tokens_out":9167,"duration_ms":80119,"temperature":0.7,"pith_summary":"This paper claims that a lightweight RGB-D salient object detector can match or exceed the accuracy of much larger models if the depth input, fusion mechanism, and feature representation are redesigned for the lightweight regime. SATNet replaces noisy raw depth maps with pseudo depth maps from a monocular depth foundation model, uses a Decoupled Attention Module for cross-modal fusion, and expands the limited feature space of a MobileNet-v2 backbone with a Dual Information Representation Module that models texture and saliency separately. On five public benchmarks, the authors report that SATNet surpasses the best lightweight methods and several heavyweight CNN methods with 5.2 million parameters and 415 FPS. The claim matters because it suggests that the usual accuracy cost of small, real-time models can be substantially reduced by addressing data quality and feature space rather than simply adding parameters.","feed_headline":"5.2M-parameter saliency network tops heavy models at 415 FPS","feed_subtitle":"Pseudo-depth maps plus dual texture-saliency features close the gap with much larger CNN baselines.","key_machinery":"The load-bearing machinery is a set of three lightweight modules built on a pair of MobileNet-v2 encoders. DAM (Decoupled Attention Module) pools RGB and depth features into horizontal and vertical vectors, projects them through fully connected layers into dual-view attention weights, and then cross-multiplies modality-specific spatial heatmaps so each branch is enhanced by the other's saliency evidence. DIRM (Dual Information Representation Module) creates two inverted pyramids, a texture pyramid supervised by edge ground truth and a saliency pyramid supervised by saliency ground truth, with two prediction heads that let gradients flow back through both; this is what expands the feature space without many extra parameters. DFAM (Dual Feature Aggregation Module) combines texture and saliency features and applies three branches of asymmetric and dilated depthwise separable convolutions to gather multi-scale receptive-field cues. The Depth Anything Model, used before training and testing, converts poor raw depth inputs into pseudo depth maps, which the paper argues removes the cross-modal inconsistency that hurts lightweight models.","core_discovery":"The paper's central claim is that the accuracy gap between lightweight and heavyweight RGB-D salient object detection is not primarily a parameter-count problem but a combination of poor depth quality, attention mechanisms unsuited to small models, and constrained feature representation. SATNet addresses each with a targeted module: Depth Anything pseudo depth maps supply smoother and more reliable geometric cues than raw depth; DAM decouples features into horizontal and vertical vectors and exchanges spatial heatmaps across modalities; DIRM builds texture and saliency feature pyramids with separate prediction heads; and DFAM merges those features using asymmetric and dilated depthwise separable convolutions that enlarge receptive fields without adding parameters. As reported on NLPR, NJU2K, SIP, STERE, and RGBD135, SATNet outperforms the convolutional heavyweight baselines and all six lightweight baselines with 5.2 M parameters, 1.5 G FLOPs, and 415 FPS.","pith_inferences":["Beyond the paper's numbers, an end-to-end deployment that generates pseudo depth on the fly would add the Depth Anything forward pass cost, which the 415 FPS figure does not include.","The pseudo-depth advantage is likely strongest when raw depth is noisy and weakest when monocular depth estimation fails, such as with transparent or textureless objects; the paper does not test those failure regimes.","DAM's horizontal and vertical decoupling is a low-rank attention proxy that could transfer to other bi-modal fusion tasks, such as RGB-thermal saliency detection, where the same speed-accuracy pressure exists.","The separate texture and saliency heads make SATNet a form of multi-task learner, suggesting that edge labels could be distilled or omitted in future variants to reduce annotation cost."],"forward_implications":["If the reported numbers hold, lightweight RGB-D salient object detection can match heavyweight CNN accuracy, making saliency detection practical on phones and embedded devices.","The pseudo-depth strategy transfers to other architectures: re-training MobileSal and AirSOD with pseudo depth improved their MAE on SIP, so depth-quality substitution is a general recipe rather than a SATNet-specific trick.","The efficiency analysis, in which DAM and DIRM account for only 14.5% of parameters and 5.4% of runtime, implies that most of the accuracy gain comes from modules that are cheap to run.","With a Swin-tiny backbone, SATNet stays competitive with transformer-based RGB-D methods at a fraction of their FLOPs, so the design generalizes beyond CNN backbones.","The same network, with EfficientNet or Swin-tiny encoders, is competitive on polyp segmentation and RGB-T salient object detection, indicating the speed-accuracy tradeoff carries over to adjacent foreground segmentation tasks."],"supporting_citations":[{"why":"Supplies the Depth Anything Model that generates the pseudo depth maps, the key data-quality fix behind the reported accuracy gains.","marker":"[5]"},{"why":"Defines the MobileNet-v2 lightweight backbone whose constrained feature space the proposed modules are designed to expand.","marker":"[30]"},{"why":"HiDANet is the heavyweight CNN baseline that SATNet claims to surpass on multiple datasets at 4.0% of its parameters and 2.1% of its FLOPs.","marker":"[9]"},{"why":"LSNet is the previous state-of-the-art lightweight method whose MAE results SATNet improves on across all five datasets.","marker":"[26]"},{"why":"AirSOD is a lightweight baseline that the paper re-trains on pseudo depth to show the depth-quality fix transfers and that SATNet still outperforms it.","marker":"[27]"},{"why":"MobileSal is an extremely efficient lightweight baseline re-trained on pseudo depth and outperformed by SATNet.","marker":"[28]"},{"why":"Channel attention is one of the heavyweight attention mechanisms replaced by DAM in the ablation and shown to fuse modalities worse in a lightweight setting.","marker":"[6]"},{"why":"Spatial and hybrid (CBAM) attention are used as baselines in the DAM ablation and visualization.","marker":"[7]"},{"why":"Self-attention is the other heavyweight attention baseline used in the DAM ablation and visualization.","marker":"[8]"},{"why":"ASPP is the receptive-field module replaced by DFAM in the ablation and outperformed in S-measure and MAE.","marker":"[31]"}],"fun_headline_variants":["Tiny 5.2M network beats heavy RGB-D saliency models at 415 FPS","Lightweight RGB-D saliency: 5.2M params, 415 FPS, top accuracy","Speed-accuracy tradeoff solved: SATNet hits 415 FPS with 5.2M params","Small model, big accuracy: RGB-D saliency with dual features","Decoupled attention and dual features make lightweight saliency SOTA"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported speed and size figures assume that generating the pseudo depth map is free or excluded, and that pseudo depth is an acceptable replacement for raw depth when comparing with raw-depth baselines.","fun_headline_variants_meta":{"raw":{"variants":["Tiny 5.2M network beats heavy RGB-D saliency models at 415 FPS","Lightweight RGB-D saliency: 5.2M params, 415 FPS, top accuracy","Speed-accuracy tradeoff solved: SATNet hits 415 FPS with 5.2M params","Small model, big accuracy: RGB-D saliency with dual features","Decoupled attention and dual features make lightweight saliency SOTA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000462,"raw_usage":{"total_tokens":2335,"prompt_tokens":996,"completion_tokens":1339,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":1226}},"tokens_in":612,"tokens_out":1339,"duration_ms":8931,"temperature":1.0,"reasoning_tokens":1226,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:21:41.426857+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the full test-time pipeline, including the Depth Anything forward pass on each RGB image, on the same GPU used to report 415 FPS; if the end-to-end throughput is materially lower, the headline speed does not describe the deployable system.","supporting_citations":[{"cited_title":"Lsnet: Lightweight spatial boosting network for detecting salient objects in rgb-thermal images,","cited_arxiv_id":null,"evidence_quote":"LSNet is the previous state-of-the-art lightweight method whose MAE results SATNet improves on across all five datasets."},{"cited_title":"Airsod: A lightweight network for rgb-d salient object detection,","cited_arxiv_id":null,"evidence_quote":"AirSOD is a lightweight baseline that the paper re-trains on pseudo depth to show the depth-quality fix transfers and that SATNet still outperforms it."},{"cited_title":"Mobilesal: Extremely efficient rgb-d salient object detection,","cited_arxiv_id":null,"evidence_quote":"MobileSal is an extremely efficient lightweight baseline re-trained on pseudo depth and outperformed by SATNet."},{"cited_title":"Attention is all you need,","cited_arxiv_id":null,"evidence_quote":"Self-attention is the other heavyweight attention baseline used in the DAM ablation and visualization."}],"review_version":1}