{"id":"55115841-e1ad-457f-a4b7-b70810ceb379","arxiv_id":"2501.04213","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"UPAQ reports large compression, speedup, and energy savings for PointPillars and SMOKE 3D detectors on an embedded GPU by jointly pruning and quantizing kernels.","lead":"UPAQ combines semi-structured pattern pruning with mixed-precision quantization to shrink 3D object detectors for autonomous vehicles. On a Jetson Orin Nano, the authors report up to 5.62x compression, 1.97x faster inference, and 2.07x lower energy versus several prior compression methods, while keeping detection accuracy at or above baseline.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The efficiency score Es (Eq. 2) that selects every kernel in Algorithms 4 and 5 depends on on-device latency and energy, but no model or measurement procedure is given, leaving the core algorithm underspecified and the Table 2 gains unattributable to the described method.","rationale":"The strongest claim is that UPAQ dominates baselines across compression, mAP, latency, and energy. For that claim to hold, the algorithm that produces the UPAQ rows must be well-defined and the measurements must be valid. The efficiency score Es in Eq. (2) is that algorithm's decision rule, but the manuscript does not specify how latency and energy are obtained for each candidate kernel. This is not a cosmetic omission: without it, Algorithms 4 and 5 are under-specified and the reported gains could come from any number of hidden scoring choices. The reader's weakest_assumption identifies the same issue, and I agree. The final on-device results are measured, but the kernels that lead to those results are selected by the undocumented Es; an independent team cannot reconstruct the method. I therefore recommend keeping the REJECT verdict. A concrete test is to have the authors provide the efficiency model and re-run the selection; if the regenerated Table 2 differs from the published table, the central claim is not supported as written.","tokens_in":13373,"tokens_out":6867,"duration_ms":69067,"concrete_test":"Request from the authors the exact efficiency model used for Latency and Energy in Eq. (2), including any calibration data, and independently re-run Algorithms 4 and 5 on PointPillars root layers with that model to regenerate the UPAQ (HCK) and UPAQ (LCK) entries of Table 2 on the Jetson Orin Nano. If the regenerated compression, mAP, latency, and energy differ from the published values beyond measurement noise, the unspecified Es is load-bearing; if the model cannot be provided, the manuscript should be revised or marked unverifiable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing gap is the efficiency score Es (Eq. 2), which is the sole selection criterion in Algorithms 4 and 5. Section IV.C states only that 'we then calculate the on-device latency and energy of the model,' but provides no formula, surrogate model, calibration, or measurement protocol for these quantities in the per-kernel search. Every pruned/quantized kernel is chosen because it maximizes Es; with Es undefined, the method is under-specified and a reader cannot run the algorithm. The final device measurements in Table 2 do not rescue this, because they are consequences of choices made by the unspecified Es: a different scoring function could select different kernels and yield different speed, energy, and mAP. In addition, the pseudocode is not executable as printed (Algorithm 4 uses temp_array before assignment; Algorithm 5 calls reshape on a Python list), reinforcing that the text lacks a well-defined method. Thus the central claim of dominating prior baselines on all axes is not supported by the manuscript's description.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UPAQ, a two-stage compression framework for 3D object detectors that combines semi-structured pattern pruning with mixed-precision quantization. It is evaluated on PointPillars and SMOKE, targeting embedded platforms such as the Jetson Orin Nano. The authors report high compression ratios (up to 5.62x), inference speedups (up to 1.97x), energy reductions (up to 2.07x), and mAP improvements over the uncompressed baselines and over four prior compression methods. The method is centered on a per-kernel efficiency score Es (Eq. 2) that is used to select pruning and quantization choices in Algorithms 4 and 5.","tokens_in":13554,"tokens_out":5690,"duration_ms":54298,"significance":"If the reported results are correct, UPAQ would be a significant advance for real-time embedded 3D detection, since it claims to simultaneously improve compression, latency, energy, and accuracy over strong baselines, including an mAP gain of about 7 points without retraining. The paper also targets 1x1 kernel compression, which is often neglected. However, the central selection mechanism is underspecified and the evaluation protocol is incomplete, so the validity of the headline claims cannot be assessed from the manuscript as written. These gaps are load-bearing for the paper's contributions and must be addressed before the work can be considered reproducible.","major_comments":[{"comment":"The efficiency score Es is the sole criterion for selecting every pruned and quantized kernel, but the manuscript never specifies how the on-device latency and energy of a candidate compressed model are computed. Section IV.C.3 states only that 'we then calculate the on-device latency and energy of the model' and insert these values into Eq. (2); no surrogate model, measurement procedure, calibration, or validation is given. Because Es directly determines every kernel selection in Algorithms 4 and 5, the reported speedups and energy reductions in Table 2 cannot be attributed to the described method. An independent implementation would have to guess the latency and energy model, making the algorithm irreproducible in its current form.","section":"Section IV.C.3, Eq. (2), Algorithms 4 and 5"},{"comment":"The evaluation protocol is not reported. The paper states a KITTI 80:10:10 split but does not specify whether this is the standard KITTI train/validation split, which object class and difficulty level are used, what IoU threshold is applied, what mAP variant is reported (e.g., 3D AP, BEV AP, R11/R40), or whether results are averaged over multiple runs. These details are essential for comparing against the cited baselines and for assessing the accuracy claims. Without them, the mAP values in Table 2 (e.g., 86.15 for PointPillars and 36.65 for SMOKE under UPAQ-LCK) cannot be verified or reproduced.","section":"Section V.A, Table 2"},{"comment":"The reported mAP improvements without retraining are anomalous and unexplained. UPAQ (LCK) improves PointPillars mAP from 78.96 to 86.15 and SMOKE from 29.85 to 36.65, while applying pruning and quantization to an already converged model with no fine-tuning. This goes against the typical behavior of post-training quantization and pruning, which generally cause accuracy loss or at most minor changes. The paper offers no mechanism for this improvement and no additional analysis, such as per-class AP, error analysis, or repeated evaluations, to rule out an evaluation artifact. This claim needs to be substantiated before the paper's dominance over prior methods can be accepted.","section":"Section V.B, Table 2"},{"comment":"The pseudocode is not executable as printed and contains undefined operations. Algorithm 4 line 6 uses temp_array before it is defined (line 3 creates temp_kernel); Algorithm 5 line 11 appends to temp_array without initialization, line 12 uses the invalid expression t1=0, and line 19 calls Flatten on a list; Algorithm 2 uses ranges written as '(0, min(n, d)]', which is not valid Python. In addition, Algorithm 4 mutates KW inside the pattern loop (line 10) without resetting temp_kernel or restoring the original weights, so each candidate is not evaluated independently and the selection is order-dependent. These issues prevent reproduction of the method and cast doubt on the correctness of the described search procedure.","section":"Algorithms 4 and 5"}],"minor_comments":[{"comment":"The term 'Compression' is not defined; please state the exact formula (e.g., original bit count divided by compressed bit count) and report per-model sparsity and bitwidth distributions so that the compression ratios can be interpreted.","section":"Table 2"},{"comment":"The paper does not specify the exact pretrained checkpoints or training configurations for PointPillars and SMOKE; please provide model sources and any required preprocessing steps.","section":"Section V.A"},{"comment":"Reference numbering is inconsistent: Ps&Qs is cited as [20] in the related work but [21] in Section V.B, and LIDAR-PTQ is cited as [22] in the related work and [23] later; please harmonize all citations.","section":"References and Section V.B"},{"comment":"The abstract says 'Jetson Orin Nano' while Section V.A says 'Jetson Orin'; please use the exact platform name consistently throughout.","section":"Abstract and Section V.A"},{"comment":"The range notation such as '(0, min(n, d)]' is ambiguous; use Python range() or standard mathematical interval notation consistently.","section":"Section IV.B, Algorithm 2"}],"recommendation":"major_revision","confidential_remarks":"The main barrier is that the method as described cannot be reproduced or verified: the efficiency model behind Es is missing, the evaluation protocol is incomplete, and the pseudocode has multiple undefined operations. These are serious but potentially fixable with additional detail and re-evaluation. If the authors can supply the missing model, correct the algorithms, and rerun the evaluation with a clearly defined protocol, the paper might be publishable; in its current form, it does not support its central claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You asked about UPAQ. The core idea is sensible: combine semi-structured pattern pruning with mixed-precision quantization for 3D detectors, and handle 1x1 kernels by grouping them into kxk blocks so they get the same pattern-based treatment as larger kernels. That 1x1 grouping is a genuine, small novelty, and the root-leaf grouping to amortize optimization across shared layers is a practical efficiency trick. The paper reports large wins on Jetson Orin for PointPillars and SMOKE—up to 5.6x compression, roughly 2x speedup, and 2x energy reduction. If those numbers hold, it is a useful engineering contribution.\n\nBut the manuscript does not back the numbers. The efficiency score Es (Eq. 2) is the sole criterion for selecting every pruned and quantized kernel in Algorithms 4 and 5, yet the latency and energy values that feed it are never defined. There is no model, no measurement procedure, no calibration. The text just says \"we then calculate the on-device latency and energy of the model.\" Without that, the algorithms are unstunnable and the reported speedups are unattributable to the described method. The pseudocode reinforces the problem: Algorithm 4 references a temp_array that is never created, and Algorithm 5 calls reshape on a Python list. Those are fixable typos, but they show the code was not checked.\n\nThe mAP claims are the other soft spot. UPAQ (LCK) improves PointPillars mAP from 78.96 to 86.15 and SMOKE from 29.85 to 36.65, with no retraining. That is a 7-point jump on KITTI. Not impossible, but with no stated split, difficulty level, IoU threshold, or mAP type, the numbers are not comparable. The 80:10:10 split is non-standard for KITTI, so the results may not align with the literature. The paper offers no mechanistic explanation for such a large accuracy gain from compression alone.\n\nThe circularity concern is real but proportionate. Es includes 1/Latency and 1/Energy, which are exactly the metrics reported as wins, so part of the gain is by construction. The fact that mAP is not in Es and still improves is the part that needs independent verification.\n\nWhat is good: the paper compares against four baselines on two platforms, covers both LiDAR and camera detectors, and the framework is described in enough detail that a motivated reader could reconstruct most of it once Es is specified. It is incremental but nontrivial, and the 1x1 grouping is worth keeping.\n\nI would send it to peer review, but with a clear ask: the authors must specify the Es model or release code, and they must report the exact KITTI protocol. If they cannot do either, it should not be accepted. The reader's REJECT is defensible for the current version; I would frame it as \"reject, but resubmission encouraged after major revision.\" It deserves referee time because the underlying idea is not silly.","headline":"UPAQ combines known pruning and quantization ideas in a practically oriented way, but the missing definition of its efficiency score and the unverifiable mAP gains make the central claims unsupported as written.","tokens_in":14173,"tokens_out":2995,"would_cite":false,"duration_ms":30114,"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":"UPAQ claims to compress 3D object detectors up to 5.62× while making them faster, more energy-efficient, and more accurate on embedded hardware.","keywords":["3D object detection","pattern pruning","mixed-precision quantization","model compression","embedded platforms","autonomous vehicles","PointPillars","SMOKE"],"falsifier":"Measure the per-kernel latency and energy of every candidate pruned-and-quantized kernel on the actual Jetson Orin Nano, recompute the efficiency score with those measured values, and verify that the kernels UPAQ selects still produce the reported 1.97× and 1.86× speedups and 2.07× and 1.87× energy reductions; if the top-scoring kernels change and the gains disappear, the central claim is falsified.","tokens_in":13095,"feed_emoji":"🚗","tokens_out":16585,"duration_ms":126814,"temperature":0.7,"pith_summary":"UPAQ is a compression framework for 3D object detectors in autonomous vehicles, combining pattern-based pruning (keeping only a few weights per kernel in diagonal, row, or column layouts) with mixed-precision quantization (different bit widths for different weights). The paper claims that on the Jetson Orin Nano embedded platform, UPAQ compresses the PointPillars LiDAR detector 5.62× and the SMOKE camera detector 5.13×, while making inference up to 1.97× and 1.86× faster and cutting energy use up to 2.07× and 1.87× compared with state-of-the-art pruning and quantization methods. In its accuracy-focused configuration, UPAQ also raises mean average precision without any retraining, from 78.96 to 86.15 on PointPillars and from 29.85 to 36.65 on SMOKE. If these numbers hold, compressed 3D perception could run in real time on the low-power hardware typically found in vehicles, addressing a major practical bottleneck for autonomous driving.","feed_headline":"5.62× smaller, 1.97× faster: UPAQ compresses PointPillars and SMOKE","feed_subtitle":"On a Jetson Orin Nano, the pruned and quantized 3D detectors also beat the original accuracy without retraining.","key_machinery":"The framework's core selection device is the efficiency score $E_s = \\alpha\\cdot\\mathrm{SQNR} + \\beta\\cdot(1/\\mathrm{Latency}) + \\gamma\\cdot(1/\\mathrm{Energy})$, which scores each candidate pruned-and-quantized kernel by combining its signal-to-quantization-noise ratio with the reciprocal of on-device latency and energy. For $k\\times k$ kernels with $k>1$, the method generates candidate sparsity patterns (main diagonal, anti-diagonal, row, or column placements of non-zero weights), quantizes the surviving weights with a symmetric mixed-precision quantizer, and keeps the pattern with the highest $E_s$; for $1\\times 1$ kernels, it flattens and regroups them into $k\\times k$ blocks so the same pattern machinery applies. A preprocessing stage groups layers into root-leaf sets by depth-first search, so the expensive search runs only on root layers and the winning pattern is replicated on coupled leaf layers. The two reported configurations, HCK (high compression) and LCK (low compression), correspond to fewer non-zero weights per kernel plus lower bit widths (e.g., 4 and 8 bits) versus more non-zero weights plus higher bit widths (e.g., 8 and 16 bits).","core_discovery":"The central claim is that a two-tier compression scheme—$k\\times k$ kernel compression with a pattern generator (main-diagonal, anti-diagonal, row, or column placements of non-zero weights) and a $1\\times 1$-to-$k\\times k$ reshape trick for point-cloud feature layers—can simultaneously shrink, speed up, and improve the accuracy of 3D object detectors. UPAQ chooses the best compression per root layer via an efficiency score that balances quantization noise against latency and energy, then propagates that pattern to coupled leaf layers. The paper reports 5.62× compression for PointPillars and 5.13× for SMOKE, up to 1.97× and 1.86× inference speedups on the Jetson Orin, up to 2.07× and 1.87× energy reduction, and higher mAP than the uncompressed baseline in the accuracy-oriented LCK variant, all with no retraining. The framework is presented as beating all prior pruning and quantization methods on every reported axis.","pith_inferences":["The paper does not test per-layer tuning of the efficiency-score weights (it fixes α=0.3, β=0.4, γ=0.3), so a natural extension is to search those weights per layer to see if the reported latency-energy trade-off can be pushed further.","The $1\\times 1$-to-$k\\times k$ reshape specifically targets point-cloud feature networks, so isolating the Pillar Feature Network in an ablation would reveal how much of the gain comes from that trick versus the root-leaf search.","The related work lists larger 3D detectors (SECOND, Focals Conv, VSC); applying UPAQ to those would test whether the compression and speedup numbers scale beyond the two architectures studied.","Since LCK improves accuracy without retraining, a short fine-tuning pass after compression might compound the gain, but the paper does not explore that combination."],"forward_implications":["If the reported numbers hold, PointPillars inference drops to 18.23 ms and SMOKE to 68.45 ms on the Jetson Orin, putting the LiDAR detector inside a real-time budget and nearly halving the camera detector's latency.","The LCK variant raises mean average precision from 78.96 to 86.15 on PointPillars and from 29.85 to 36.65 on SMOKE without retraining, showing that pruning and quantization need not cost accuracy.","Because the expensive search runs only on root layers and the chosen pattern is copied to leaf layers, compression cost scales with the number of root layers rather than the total layer count, which matters for deeper detectors.","Both a LiDAR-based detector and a camera-based detector improve, so the method is not tied to one input modality among 3D perception pipelines.","The HCK configuration gives users a compression-first option and the LCK configuration an accuracy-first option, a tunable trade-off controlled by the number of non-zero weights per kernel and the quantization bit widths."],"supporting_citations":[{"why":"Defines the PointPillars LiDAR detector architecture that UPAQ compresses and evaluates.","marker":"[14]"},{"why":"Defines the SMOKE camera-based detector architecture that UPAQ compresses and evaluates.","marker":"[16]"},{"why":"Supplies the Ps&Qs quantization-aware pruning baseline that UPAQ compares against.","marker":"[20]"},{"why":"Supplies the CLIP-Q pruning-quantization baseline that UPAQ compares against.","marker":"[21]"},{"why":"Supplies the LIDAR-PTQ post-training quantization baseline for point-cloud detectors.","marker":"[22]"},{"why":"Supplies the R-TOSS entry-pattern semi-structured pruning baseline that UPAQ directly extends.","marker":"[23]"},{"why":"Supplies the inference acceleration framework used for deployment and latency measurements.","marker":"[29]"},{"why":"Supplies the KITTI dataset on which all accuracy, latency, and energy numbers are measured.","marker":"[33]"}],"fun_headline_variants":["UPAQ: 5.6× smaller, 1.9× faster, more accurate, no retraining","UPAQ's pruning+quantization makes 3D detectors 5.6× smaller, 1.9× faster","UPAQ: 5.6× compression, 1.9× speedup, 2× energy cut on Jetson Orin","Without retraining, UPAQ beats PointPillars and SMOKE in size, speed, and mAP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the efficiency score's latency and energy terms come from a reliable model of on-device performance; the paper never specifies how those values are measured or predicted, so if that model is wrong, the reported speedups and energy savings may be an artifact of the scoring function rather than the compression.","fun_headline_variants_meta":{"raw":{"variants":["UPAQ: 5.6× smaller, 1.9× faster, more accurate, no retraining","UPAQ's pruning+quantization makes 3D detectors 5.6× smaller, 1.9× faster","UPAQ: 5.6× compression, 1.9× speedup, 2× energy cut on Jetson Orin","Without retraining, UPAQ beats PointPillars and SMOKE in size, speed, and mAP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001824,"raw_usage":{"total_tokens":7171,"prompt_tokens":937,"completion_tokens":6234,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":6108}},"tokens_in":553,"tokens_out":6234,"duration_ms":43495,"temperature":1.0,"reasoning_tokens":6108,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:39:15.765824+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the per-kernel latency and energy of every candidate pruned-and-quantized kernel on the actual Jetson Orin Nano, recompute the efficiency score with those measured values, and verify that the kernels UPAQ selects still produce the reported 1.97× and 1.86× speedups and 2.07× and 1.87× energy reductions; if the top-scoring kernels change and the gains disappear, the central claim is falsified.","supporting_citations":[{"cited_title":"Pointpillars: Fast encoders for object detection from point clouds","cited_arxiv_id":null,"evidence_quote":"Defines the PointPillars LiDAR detector architecture that UPAQ compresses and evaluates."},{"cited_title":"Smoke: Single-stage monocular 3d object detection via keypoint estimation","cited_arxiv_id":null,"evidence_quote":"Defines the SMOKE camera-based detector architecture that UPAQ compresses and evaluates."},{"cited_title":"Ps and qs : Quantization-aware pruning for efficient low latency neural network inference","cited_arxiv_id":null,"evidence_quote":"Supplies the Ps&Qs quantization-aware pruning baseline that UPAQ compares against."},{"cited_title":"Clip-q: Deep network compression learning by in - parallel pruning -quantization","cited_arxiv_id":null,"evidence_quote":"Supplies the CLIP-Q pruning-quantization baseline that UPAQ compares against."},{"cited_title":"R-TOSS: A framework for real-time object detection using semi -structured pruning","cited_arxiv_id":null,"evidence_quote":"Supplies the R-TOSS entry-pattern semi-structured pruning baseline that UPAQ directly extends."},{"cited_title":"TensorRT-Based Framework and Optimization Methodology for Deep Learning Inference on Jetson Boards","cited_arxiv_id":null,"evidence_quote":"Supplies the inference acceleration framework used for deployment and latency measurements."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the KITTI dataset on which all accuracy, latency, and energy numbers are measured."}],"review_version":1}