{"id":"1349494c-3c3f-40cf-8770-4c9189de6b6f","arxiv_id":"1908.07842","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MobileNet-V2 with mixed precision training achieves 27.77 fps on NVIDIA Xavier with about 6 points lower CMC-1 than ResNet-50 single precision on three re-ID benchmarks.","lead":"A lightweight person re-identification model trained with half precision runs at 27.77 frames per second on an NVIDIA Xavier edge device. It is faster and more power-efficient than a ResNet-50 baseline, at a moderate accuracy cost.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"3.25x/1.45x headline gains conflate MobileNet-V2 architecture, FP16, and DLA-vs-Volta hardware; same-model columns show precision-only gains of 1.38x throughput and 1.0x power.","rationale":"The reader's weakest-assumption analysis is exactly the load-bearing concern: the headline speedups and power savings are not a clean measurement of mixed precision. The paper discloses that half precision maps to DLA and single precision to Volta, and its own per-same-model columns reduce the MobileNet-V2 precision-only improvement to 1.38x throughput and 1.0x power. However, the paper does provide a useful system-level comparison with reproducible code and the key confound is at least partially disclosed in Tables 6-7, so this does not justify rejection. The accuracy discrepancy is secondary but real. Because the central engineering result remains plausible at the system level, the appropriate action is to require the authors to recompute/reframe the headline claims, which is exactly the conditional acceptance the reader recommended. Thus my verdict is unchanged.","tokens_in":8794,"tokens_out":5482,"duration_ms":50988,"concrete_test":"Run the authors' released code on the same NVIDIA Xavier with compute unit held fixed: (1) MobileNet-V2 FP32 on Volta, (2) MobileNet-V2 FP16 on Volta with DLA disabled, (3) MobileNet-V2 FP16 on DLA, (4) ResNet-50 FP32 on Volta. If FP16-on-Volta versus FP32-on-Volta gives about 1.38x throughput and 1.0x power, the abstract's 3.25x/1.45x claims are confirmed to be architecture-plus-hardware gains, not mixed-precision gains. Separately, recompute the average CMC-1 gap from Fig. 5; whether it is 5.6% or 6.1 percentage points should be resolved and the abstract corrected accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract attributes the headline gains to the \"MobileNet-V2 mixed precision training method.\" However, Section 4.4 states that half-precision models run on the Xavier DLAs while single-precision models run on the Volta GPU cores, and Tables 6-7's \"Over the Baseline\" columns compare MobileNet-V2 FP16 on DLA to ResNet-50 FP32 on Volta. The \"Per Same Model\" columns, which hold the architecture constant, show only 1.38x throughput and 1.0x power for MobileNet-V2, and even ResNet-50's 2.54x is DLA-vs-GPU rather than pure precision. Thus the 3.25x and 1.45x numbers are system-level gains from changing backbone, precision, and compute unit simultaneously; the abstract's wording misattributes them to mixed precision. The underlying experiment is disclosed, so the engineering comparison remains useful, but the marginal contribution of mixed precision is not measured. In addition, the abstract's 5.6% accuracy deterioration is not directly supported by Fig. 5: MobileNet-V2 mixed CMC-1 mean is 70.3 versus ResNet-50 single 76.4, a 6.1 percentage-point gap; the origin of 5.6% should be stated or the figure corrected.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript reports an empirical study of person re-identification at the edge. The authors fine-tune ResNet-50 and MobileNet-V2 with a triplet loss using either single precision or mixed precision, evaluate on CUHK03, Market-1501, and DukeMTMC, and measure training time, model size, inference throughput, and power on an NVIDIA Xavier. They report that MobileNet-V2 with mixed precision reaches 27.77 fps at 6.48 W with an average accuracy loss of 5.6% relative to the ResNet-50 single-precision baseline, and claim 3.25x throughput, 1.75x training-time, and 1.45x power improvements. The paper includes a qualitative comparison and the authors state that code and pre-trained networks are publicly available.","tokens_in":9089,"tokens_out":8574,"duration_ms":73792,"significance":"The paper addresses a relevant deployment problem: real-time person re-ID on constrained hardware. Its strengths are the reproducible artifacts (public code/models), realistic measurements on a current edge platform, and an empirical comparison on three standard benchmarks. If the stated contributions are re-framed carefully, the system-level comparison is a useful engineering data point for practitioners. The claimed methodological finding—that batch-normalization inputs and the triplet loss must remain in FP32 during mixed-precision training—is interesting but currently lacks ablation support. The main limitation is that the headline speed, power, and training-time numbers are presented as effects of mixed precision even though the experiments change the backbone, precision, and compute unit simultaneously.","major_comments":[{"comment":"The headline improvements are not cleanly attributable to mixed precision. Section 4.4 states that half-precision models run on the Xavier DLAs while single-precision models run on the Volta GPU cores. Consequently, Table 6's 'Over the Baseline' figure of 3.25x compares MobileNet-V2 FP16 on DLA with ResNet-50 FP32 on Volta, and Table 7's 1.45x power improvement is the ratio of ResNet-50 FP32 power to MobileNet-V2 FP16 power. Even the 'Per Same Model' columns (1.38x throughput, 1.0x power for MobileNet-V2) are not precision-only comparisons because the mixed-precision rows use DLA and the single-precision rows use Volta. Likewise, the 1.75x training-time speedup in Table 3 is the ratio of ResNet-50 single-precision time (242.65 min) to MobileNet-V2 mixed-precision time (138.1 min), while the same-model speedup for MobileNet-V2 is only 140.3/138.1 = 1.016x. The abstract therefore overstates what is measured. Please either add a same-compute-unit precision comparison (e.g., FP16 vs FP32 on the Volta GPU) or rephrase the claims as system-level gains from changing the backbone, precision, and compute engine together.","section":"Section 4.4 (Tables 6-7), Section 4.3 (Table 3), Abstract"},{"comment":"The claim that MobileNet-V2 mixed precision 'deteriorates accuracy only 5.6%' cannot be derived from the reported results. Figure 5 gives mean CMC-1 values of 76.4 for ResNet-50 single precision and 70.3 for MobileNet-V2 mixed precision, a 6.1 percentage-point gap (approximately 8% relative), not 5.6%. No other metric in Figures 3-5 yields 5.6% as an average over the three datasets. Please state the exact metric and computation behind the 5.6% figure or correct the abstract and Section 4.2. In addition, no variance or number of repeats is reported, so the 'negligible' 0.5-point difference between MobileNet-V2 single and mixed precision is not established as meaningful.","section":"Abstract, Section 4.2, Figure 5"},{"comment":"The key empirical rule that 'inputs of batch normalization layers and the loss calculation should be in FP32' is stated as a finding, but no experiment supports it. The reader cannot tell whether FP16 BN inputs actually prevent convergence, whether FP16 loss harms hard-sample mining, or whether these choices merely have a small effect. Since this is one of the paper's stated contributions and is used to justify the mixed-precision design, please include an ablation or training curves with and without the FP32 constraint.","section":"Section 3.3"}],"minor_comments":[{"comment":"The citation 'Andrew et al.[4]' should refer to the MobileNets paper (reference [10] in the bibliography), not to the ImageNet reference [4].","section":"Section 3.1.2"},{"comment":"There are typos: 'relive the pressure' should be 'relieve the pressure' in Section 1, and 'concider' should be 'consider' in Section 3.1.1.","section":"Section 1 and Section 3.1.1"},{"comment":"The sentence 'single precision negligibility deteriorate the CMC-1 performance for 0.5%' is ungrammatical and seems to state the opposite of the data; please rewrite to say that mixed precision is 0.5 percentage points below single precision.","section":"Section 4.2"},{"comment":"The in-bar labels in Figures 3-5 are very small, and the exact mean values matter for the paper's accuracy claims; please also provide the mean values in a table.","section":"Figures 3-5"},{"comment":"The column headings of Table 5 are confusing: it is not immediately clear whether 'Per Same Model' and 'Over the Baseline' apply to both precision columns. Please reformat the table so the improvement factors are unambiguous.","section":"Table 5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an engineering/experimental paper; the experiments are disclosed, and the code and models appear to be available. The main issue is that the abstract and conclusion present system-level improvements as if they were attributable to mixed precision alone. I do not see a novelty-disclosure problem, but the authors should be asked to recalibrate all headline numbers and, if feasible, to add a same-hardware FP16/FP32 comparison. The paper would be acceptable for a venue of this type after those revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my take on Baharani et al. The paper earns a read from anyone deploying person re-ID on edge hardware. The genuinely new, transferable result is the empirical finding that mixed-precision training for re-ID should keep batch norm inputs and the triplet loss in FP32, or training diverges. They also show MobileNet-V2 can run re-ID at about 27.8 fps at 6.5 W on a Jetson Xavier, which is practically useful. Credit where due: the evaluation is on standard benchmarks (CUHK03, Market-1501, DukeMTMC), the setup is disclosed in Section 4.4, and code and pretrained models are public.\n\nNow the soft spots. The abstract attributes a 3.25x throughput gain and 1.45x power saving to \"MobileNet-V2 mixed precision training.\" But the experiment changes three variables at once: backbone (ResNet-50 to MobileNet-V2), precision (FP32 to FP16), and compute unit (Volta GPU cores to the deep-learning accelerators). The paper's own per-same-model columns show the precision-only gain is 1.38x throughput and 1.0x power for MobileNet-V2. The headline numbers are system-level gains, not precision gains. That's disclosed in the tables but not in the abstract, and the distinction matters for anyone deciding whether mixed precision alone is worth the engineering effort.\n\nSecond, the 5.6% accuracy deterioration claim is not consistent with Fig. 5. That figure gives CMC-1 means of 76.4 for ResNet-50 single precision and 70.3 for MobileNet-V2 mixed precision: a 6.1 percentage-point drop, about 8% relative, with a larger gap on CUHK03. The authors need to either derive the 5.6% number or correct the figure. Also, there are no error bars anywhere, so it is hard to know which gaps are real versus noise.\n\nIs the central argument sound? I think the engineering conclusion holds: a light-weight backbone plus mixed precision gives real-time, low-power re-ID with acceptable accuracy loss. The paper does not overclaim the science; it is an application paper. My recommendation: send it to a serious referee. The headline and abstract need to be rewritten to separate system-level gains from precision-only gains, the accuracy discrepancy needs to be resolved, and the evaluation would benefit from variance information. But the underlying measurements are reproducible and the BN/loss precision finding is a useful data point.","headline":"Useful edge-deployment study, but the headline speedup/power numbers conflate architecture, precision, and hardware, and the accuracy claim doesn't match the paper's own figures.","tokens_in":9620,"tokens_out":2354,"would_cite":false,"duration_ms":23464,"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 person re-identification system built on MobileNet-V2 and trained in mixed precision achieves real-time edge inference at 27.77 fps and 6.48 W, with only a 5.6% average rank-1 accuracy loss over a ResNet-50 single-precision baseline.","keywords":["person re-identification","mixed precision training","edge computing","MobileNet-V2","ResNet-50","triplet loss","NVIDIA Xavier","real-time inference"],"falsifier":"Run the identical MobileNet-V2 model in FP32 and FP16 on the same processor (both on Volta GPU cores, or both on the Deep Learning Accelerator) and measure throughput, power, and accuracy; if FP16 does not improve throughput or cut power, the paper's precision-driven gains collapse.","tokens_in":8574,"feed_emoji":"📹","tokens_out":7595,"duration_ms":70333,"temperature":0.7,"pith_summary":"The paper argues that real-time person re-identification—recognizing the same person across different camera views—does not require a heavy server-side network. It shows that a MobileNet-V2 backbone trained with mixed FP32/FP16 precision and deployed in half precision on an NVIDIA Xavier edge node identifies people at 27.77 frames per second while consuming 6.48 W. Against a ResNet-50 single-precision baseline on CUHK03, Market-1501, and DukeMTMC, this is a 3.25x throughput gain, a 1.45x power reduction, and 1.75x faster training, at an average accuracy cost of 5.6%. The paper also pins down which parts of a re-ID network tolerate half precision: convolutions and matrix multiplies can run in FP16, but batch-normalization inputs and the triplet loss must stay in FP32 for training to converge and hard-sample mining to work.","feed_headline":"Real-time person re-ID at the edge: 27.77 fps, 6.48 W","feed_subtitle":"FP16 MobileNet-V2 cuts training time 1.75x and power 1.45x, losing only 5.6% accuracy.","key_machinery":"The central mechanism is mixed-precision training configured by a specific layer partition: convolution and general matrix multiply (GeMM) operations are assigned to half precision (FP16), while batch-normalization inputs and the triplet-loss calculation remain in single precision (FP32). The paper reports that FP16 batch-norm inputs stop training from converging, and an FP16 loss computation degrades the hard-sample pool used for triplet mining. The MobileNet-V2 backbone provides the light-weight representation, the triplet loss with hard mining provides the re-identification objective, and on the NVIDIA Xavier edge node the FP16 models run on the Deep Learning Accelerators while FP32 models run on the Volta GPU cores.","core_discovery":"The paper claims that mixed precision is a usable deployment lever for person re-ID at the edge, not just a training-time trick. Concretely, replacing the ResNet-50 backbone with MobileNet-V2 and training in mixed precision lets inference run at 27.77 fps and 6.48 W on NVIDIA Xavier while keeping rank-1 accuracy within 5.6% of the ResNet-50 single-precision baseline on three benchmarks. The authors further report that half precision alone degrades ResNet-50 by only 0.9% on average, and that the FP16 MobileNet-V2 model is 18.92x smaller than the baseline. The central engineering rule that emerges is a layer-wise precision partition: convolution and GeMM operations go to FP16, while batch-normalization inputs and the triplet-loss computation stay in FP32, because FP16 in those places prevents convergence and weakens the hard-positive/negative mining that triplet loss depends on.","pith_inferences":["The paper's tables imply a distinction its abstract does not make explicit: the headline 3.25x throughput and 1.45x power gains compare MobileNet-V2 FP16 on the Deep Learning Accelerator against ResNet-50 FP32 on the Volta GPU cores, so the bundled gain includes the backbone swap and the compute-unit change, not just precision.","A clean test would hold the compute unit fixed and compare FP32 versus FP16 for the same network; the paper does not report that comparison, so the pure precision contribution to throughput and power remains untested.","Because the accuracy cost of half precision for ResNet-50 is much smaller than the cost of switching to MobileNet-V2, most of the measured speedup likely comes from the architecture swap rather than from FP16 arithmetic.","The reported 27.77 fps covers feature extraction only, not detection or tracking; an end-to-end multi-camera tracking pipeline would consume a larger share of the edge node's time and power budget."],"forward_implications":["At 27.77 fps, the FP16 MobileNet-V2 model clears the common 25 fps real-time video bar, so a single edge node can keep pace with live camera streams.","The precision-partitioning rule—conv and GeMM in FP16, batch-norm inputs and loss in FP32—is presented as transferable to other re-ID networks, allowing existing high-accuracy designs to move to the edge without architectural changes.","The model size drops from 94.6 MB for ResNet-50 single precision to 5.0 MB for FP16 MobileNet-V2, easing storage and memory-bandwidth constraints on embedded hardware.","Training time falls roughly from 242.65 to 138.1 minutes for the MobileNet-V2 comparison, so mixed precision also shortens the server-side tuning loop.","Because half precision alone costs ResNet-50 only about 0.9% average accuracy, precision reduction appears to be a cheaper deployment knob than backbone substitution."],"supporting_citations":[{"why":"Defines the ResNet-50 residual network used as the baseline backbone for accuracy, throughput, and power comparisons.","marker":"[7]"},{"why":"Supplies the triplet-loss formulation with hard sample mining that all four network variants are trained with.","marker":"[8]"},{"why":"Introduces the depthwise separable convolution design that the light-weight MobileNet-V2 backbone builds on.","marker":"[10]"},{"why":"Provides the mixed-precision training mechanism (FP32 master weight copies and gradient scaling) that the paper partitions across network layers.","marker":"[14]"},{"why":"One of the three re-ID benchmarks (CUHK03) used to measure accuracy.","marker":"[13]"},{"why":"One of the three re-ID benchmarks (Market-1501) used to measure accuracy.","marker":"[24]"},{"why":"One of the three re-ID benchmarks (DukeMTMC) used to measure accuracy.","marker":"[16]"},{"why":"The re-ranking method that improves mAP in the reported accuracy results.","marker":"[26]"}],"fun_headline_variants":["Mixed precision re-ID: 27.77 fps at 6.48 W, 5.6% accuracy loss","FP16 + MobileNet-V2: edge re-ID 3.25x faster, 1.45x less power","Edge re-ID with FP16: 27.77 fps, 6.48 W, 18.92x smaller model","Re-ID at edge: mixed precision hits 27.77 fps, 6.48 W"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline speed and power numbers compare two networks running on two different processors at the same time, so they are not a clean measurement of what mixed precision alone buys.","fun_headline_variants_meta":{"raw":{"variants":["Mixed precision re-ID: 27.77 fps at 6.48 W, 5.6% accuracy loss","FP16 + MobileNet-V2: edge re-ID 3.25x faster, 1.45x less power","Edge re-ID with FP16: 27.77 fps, 6.48 W, 18.92x smaller model","Re-ID at edge: mixed precision hits 27.77 fps, 6.48 W"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001178,"raw_usage":{"total_tokens":4896,"prompt_tokens":1002,"completion_tokens":3894,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":3777}},"tokens_in":618,"tokens_out":3894,"duration_ms":26562,"temperature":1.0,"reasoning_tokens":3777,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:26:28.004137+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical MobileNet-V2 model in FP32 and FP16 on the same processor (both on Volta GPU cores, or both on the Deep Learning Accelerator) and measure throughput, power, and accuracy; if FP16 does not improve throughput or cut power, the paper's precision-driven gains collapse.","supporting_citations":[{"cited_title":"In: International Conference on Learning Representations (2018)","cited_arxiv_id":null,"evidence_quote":"Provides the mixed-precision training mechanism (FP32 master weight copies and gradient scaling) that the paper partitions across network layers."},{"cited_title":"In: CVPR (2014)","cited_arxiv_id":null,"evidence_quote":"One of the three re-ID benchmarks (CUHK03) used to measure accuracy."},{"cited_title":"In: Computer Vision, IEEE International Conference on (2015)","cited_arxiv_id":null,"evidence_quote":"One of the three re-ID benchmarks (Market-1501) used to measure accuracy."},{"cited_title":"In: European Conference on Computer Vision workshop on Benchmarking Multi-Target Tracking (2016)","cited_arxiv_id":null,"evidence_quote":"One of the three re-ID benchmarks (DukeMTMC) used to measure accuracy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The re-ranking method that improves mAP in the reported accuracy results."}],"review_version":1}