{"id":"8c0ebc6e-8c89-4ec3-a656-422b61a63321","arxiv_id":"2504.14471","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"The proposed PICO/LeAFNet pipeline compresses point clouds by fitting two small neural networks with learnable activations, reporting 4.92 dB average BD-PSNR gain over MPEG codecs on 8iVFB.","lead":"A point cloud compression method encodes geometry and colors into two small neural networks whose weights are quantized and entropy-coded, then decodes by querying the networks. It reports large bitrate savings over the MPEG G-PCC and V-PCC standards on a four-sequence benchmark.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Decoder cannot reconstruct the geometry: Algorithm 1 derives the sampling space V from the original point cloud X but never transmits V or W, so the reported bpp omits mandatory side information.","rationale":"The reader correctly identified the bitrate accounting as load-bearing, but the specific item they named (tau_0 and model dictionary selection) is small relative to the real problem. Algorithm 1 computes V from the original point cloud and never sends it, so the decoder cannot even define the coordinate set to query. This is an internal inconsistency that the paper's own pseudocode exposes. The concern is not that the method is necessarily false; it is that the headline 4.92 dB gain is computed without counting mandatory side information, so the claim cannot be taken at face value. A corrected paper could add the side-information cost and either confirm or reduce the gain. That is exactly the condition the reader's verdict already imposes, so I keep the verdict unchanged rather than escalate to rejection. The paper still has merit: the two-stage decomposition, the LeAFNet ablation, and the unimodality study are plausible contributions, but the missing V/W signaling is a concrete, checkable gap that must be resolved before the quantitative claims are accepted.","tokens_in":15630,"tokens_out":4429,"duration_ms":42236,"concrete_test":"Reproduce the codec as described and add the missing side information: for each sequence (start with longdress), count the non-empty M=5 cubes, entropy-code that occupancy map (e.g., a compressed bitmask or octree over the 32^3 cube grid), add the resulting bits to PICO's bpp for every RD point in Table 1, and recompute BD-PSNR and BD-BR against G-PCC. Also confirm that Algorithm 1's decode stage cannot produce Xhat when only the network parameters are decoded. Report the new average gain; if it drops materially or the decoder fails, the 4.92 dB claim is not supported as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing gap is not the threshold tau_0 but the sampling space V itself. In Section 3.2.2, W and V are defined from the original geometry X (Eqs. 11-12), and Algorithm 1 uses V at both encode (lines 2, 8, 9, 11) and decode (lines 19-20). However, line 17 entropy-encodes only the quantized network parameters; V (equivalently W, the set of non-empty M=5 cubes) is never transmitted. The decoder therefore cannot evaluate f_g on the correct coordinate set: it does not know which 32x32x32 cubes contain points. Reconstructing by querying all of S is not the described algorithm and is far more expensive, and f_g was not trained on voxels outside V. So W is mandatory side information, and its bit cost is absent from every bpp number in Table 1 and Figure 2. This directly affects the headline 4.92 dB BD-PSNR gain: the gain is computed against a bitrate that omits a component of the bitstream. The same applies to tau_0 and model selection, but those are small scalars; W is a 32768-cube occupancy pattern (before compression) and can dominate at low bpp. This is an internal inconsistency in the described codec, not merely a disagreement with current consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PICO, an implicit neural representation (INR) framework for static point cloud compression, decomposed into two stages: geometry compression and attribute compression. The geometry stage trains a neural network to predict voxel occupancy probabilities over a restricted sampling space, applies a dynamically optimized threshold, and the attribute stage regresses colors on the reconstructed geometry. The network architecture, LeAFNet, replaces conventional MLP activations with learnable radial-basis-function (RBF) inspired activations in the KAN style. The network parameters are quantized and entropy-coded with DeepCABAC, framing compression as neural parameter compression. Experiments on the first frames of four 8iVFB sequences report large BD-BR savings and BD-PSNR/BD-PCQM gains over G-PCC (octree), G-PCC (trisoup), and V-PCC, including the headline average 4.92 dB D1 PSNR improvement.","tokens_in":15880,"tokens_out":6649,"duration_ms":57204,"significance":"If the reported bitrate accounting is correct, PICO would be a noteworthy INR-based codec that beats the MPEG standards on this benchmark with a parameter-efficient KAN-inspired architecture. The paper includes a detailed algorithm, several ablations (threshold selection, sampling, positional encoding, model selection), and public baselines. The main strengths are the clean two-stage formulation, the demonstration of LeAFNet's parameter efficiency over an MLP, and the effort to control rate via L1 regularization and entropy coding. However, the central claim depends critically on the completeness of the transmitted bitstream, and the experimental validation is narrow. As written, the bitrate accounting has a load-bearing gap that must be resolved before the headline results can be accepted.","major_comments":[{"comment":"The decoder in Algorithm 1 reconstructs X at lines 19-20 by querying f_g and f_a over the sampling space V, but V (and its defining cube occupancy set W) is computed from the original point cloud X in Eqs. (11)-(12) and is never included in the bitstream: line 17 entropy-encodes only the quantized network parameters. The decoder therefore cannot determine which of the 2^{3M}=32768 cubes (for M=5) contain points, so W is mandatory side information whose bitrate is absent from every bpp value in Table 1 and Figure 2. This omission can be large at low bpp and directly inflates the reported BD-BR savings and the headline 4.92 dB average D1 PSNR gain. The scalar threshold tau_0 from Algorithm 1 line 10 and the model dictionary index from SelectModel (line 3) are also not specified as transmitted. The paper must report the full bitstream including all side information, or explain how the decoder regenerates V without ground-truth geometry, and include that cost in all rate values.","section":"Algorithm 1, Section 3.2.2, Eqs. (11)-(12)"},{"comment":"The rate-distortion experiments are not reproducible as reported. Section 4.1.3 gives training steps, quantization steps, and optimizer settings, but does not give the L1 regularization strengths lambda_g and lambda_a used for each RD point, the entries of the model dictionary (hidden dimensions, number of RBF centers), the RBF bandwidth h, or the number and bpp values of the RD points in Figure 2. Section 3.5.2 states that a model dictionary M enables adaptive selection by target bpp but never lists the dictionary. Without these parameters, the RD curves in Figure 2 and the BD metrics in Table 1 cannot be verified, weakening the claim that PICO achieves precise rate control through joint tuning of lambda and Delta.","section":"Section 4.1.3, Section 3.5.2"},{"comment":"The experimental evaluation is limited to the first frame of each of the four 8iVFB sequences, with no error bars or repeated trials. The paper motivates the method with LiDAR-acquired large-scale point clouds in the Introduction, but all results are on a single dataset of human bodies at 1024^3 resolution. The absence of multi-frame statistics and the absence of any large-scale outdoor dataset make the broad claims of 'superior compression performance' and 'state-of-the-art' insufficiently supported, particularly since the method's complexity (training an INR per point cloud) is not evaluated for scalability.","section":"Section 4.2, Section 1"}],"minor_comments":[{"comment":"The caption 'Static PCC Performance of PICO (MLP), G-PCC (octree), G-PCC (trisoup) and V-PCC against PICO' is ambiguous. It should clarify that each column pair reports the BD gains of the proposed PICO (LeAFNet) relative to the named baseline, with the first pair being the LeAFNet-versus-MLP ablation.","section":"Table 1 caption"},{"comment":"In Eq. (22), the two directional errors are both written as e(B,A); one should be e(A,B) to define the symmetric point-to-point error correctly.","section":"Eq. (22)"},{"comment":"The sampling strategy notation in Eqs. (13)-(14) is confusing: Eq. (13) uses V-X while Eq. (14) uses U(V), and the relationship between the calibrated rate alpha-hat and the original alpha is not explained in the main text.","section":"Eqs. (13)-(14)"},{"comment":"The unimodality experiment in Figure 3(b) should specify the sequence, model configuration, and the range of thresholds used, since only the lambda values are currently identifiable.","section":"Figure 3(b)"},{"comment":"The paper cites Ruan et al. [32] as prior work on implicit neural compression of point clouds but does not compare against it in the experiments; adding this baseline (or justifying its exclusion) would strengthen the state-of-the-art claim.","section":"Section 2.2, References"}],"recommendation":"major_revision","confidential_remarks":"The missing side-information cost for the sampling space W is the central concern. If the authors can show that W can be transmitted at negligible cost, or that the decoder can infer it without ground-truth geometry, the result may survive; otherwise the headline gains may be substantially reduced. I recommend the editor request the authors provide a complete bitstream specification and recompute all rate-distortion metrics, and also expand the evaluation to at least full sequences or multiple frames. The paper's narrow evaluation (one frame per sequence) is also a concern for a venue claiming broad compression improvements."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline number – 4.92 dB BD-PSNR over MPEG anchors – is not load-bearing as written. The decoder cannot reconstruct the geometry it promises. In Algorithm 1, the sampling space V is defined from the original point cloud X via the occupied cube set W, and both encode and decode steps use V. But the bitstream contains only the compressed network parameters (line 17). Nothing tells the decoder which 32x32x32 cubes are occupied. Without W (or an equivalent signal), the decoder does not know where to evaluate f_g, and f_g was never trained on voxels outside V. Querying all of S is not the described algorithm and is far more expensive; it also may produce arbitrary outputs outside the training domain. The same goes for tau_0 and the model dictionary selection, though those are small scalars. W is not small: it is a 32768-cube occupancy pattern. So every bpp in Table 1 and Figure 2 omits a mandatory part of the bitstream, and the BD gains are computed against understated bitrates.\n\nThis is a fixable problem in principle – transmit W as side information – but as written the codec is internally inconsistent. That is the main thing you should know.\n\nWhat the paper does well: the LeAFNet activation using RBFs is a sensible, KAN-inspired modification, and the ablations are thoughtful. The two-stage geometry/attribute split is a reasonable way to handle class imbalance, and the dynamic threshold search is a nice touch. If the rate accounting were fixed and the method still beat the MPEG anchors, this would be an interesting contribution to INR-based point cloud compression.\n\nBut there are other soft spots. Only the first frame of each 8iVFB sequence is used, there are no error bars or learned baselines, no code or data is released, and hyperparameters like lambda values and exact model sizes for each RD point are absent. None of that is disqualifying by itself; the V/W omission is.\n\nWho is this for? A practitioner in learned point cloud compression who wants to track INR codec progress. It deserves a serious referee because the methods and results are plausible enough to warrant scrutiny, but the paper needs major revision and explicit bitstream accounting before any of the headline numbers can be taken at face value. I would send it to review, and I would tell the authors exactly what I wrote above.","headline":"The 4.92 dB BD-PSNR claim is not supportable as written: the decoder is never told which cubes to evaluate, so the reported bpp omits mandatory side information.","tokens_in":16487,"tokens_out":3643,"would_cite":false,"duration_ms":33317,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"PICO, an INR codec with learnable activations, outperforms MPEG point cloud compression by 4.92 dB on geometry.","keywords":["point cloud compression","implicit neural representation","learnable activation function","Kolmogorov-Arnold networks","geometry compression","attribute compression","neural parameter compression","entropy coding"],"falsifier":"Recompute the rate-distortion curves for the four 8iVFB sequences while charging bits for the dynamic threshold (one scalar per sequence) and the model dictionary index (a few bits per rate point); if the BD-PSNR gain over G-PCC (octree) drops below the reported 7.09 dB average, the central claim fails. Alternatively, run the decoder with a fixed threshold derived only from the transmitted model and measure the resulting D1 PSNR loss.","tokens_in":15344,"feed_emoji":"🧊","tokens_out":8510,"duration_ms":65385,"temperature":0.7,"pith_summary":"The paper proposes PICO, a point cloud codec built on implicit neural representations (INRs), and reframes compression as the task of compressing the network's parameters. It splits the problem into geometry compression, posed as voxel-occupancy classification, and attribute compression, posed as color regression on the reconstructed surface. The central claim is that PICO's geometry compression beats the MPEG standards G-PCC (octree and trisoup) and V-PCC, with an average gain of 4.92 dB in D1 PSNR and about 53.5% bitrate reduction on the 8iVFB dataset, and that joint geometry-attribute compression is competitive. The gain is attributed to LeAFNet, a network whose learnable activation functions fit implicit fields with fewer parameters than conventional MLPs.","feed_headline":"INR point cloud codec beats MPEG standard by 4.92 dB","feed_subtitle":"PICO turns geometry into network weights and outperforms G-PCC and V-PCC on 8iVFB.","key_machinery":"LeAFNet is the central object: a coordinate network whose hidden layers use trainable activation functions of the form $\\phi(x) = w_b\\,\\mathrm{silu}(x) + w_s \\sum_i \\exp(-\\|x-c_i\\|^2/h^2)$, replacing the B-spline kernels of Kolmogorov-Arnold Networks with radial basis functions. The argument also depends on three supporting mechanisms: a dynamic threshold $\\tau_0$ chosen by golden-section search over a D1-PSNR curve that is empirically unimodal, which decides which voxels count as occupied; a reweighted sampling strategy that keeps occupied and empty voxels balanced despite severe sparsity; and parameter compression by uniform quantization followed by DeepCABAC entropy coding, with rate controlled by $\\ell^1$ regularization and a dictionary of model sizes.","core_discovery":"The discovery is that a per-point-cloud optimized network with learnable radial-basis activations can, after quantization and entropy coding of its parameters, reconstruct voxelized geometry at higher D1 PSNR than the standardized G-PCC and V-PCC codecs at the same bitrate on the four 8iVFB test sequences. The authors report BD-PSNR gains of 7.09 dB over G-PCC (octree), 3.95 dB over G-PCC (trisoup), and 3.73 dB over V-PCC, and an average 53.54% BD-BR reduction for geometry. For joint geometry and attribute compression they report a BD-PCQM improvement of $2.70 \\times 10^{-3}$ over the same anchors, which they take as evidence that the two-stage decomposition preserves perceptual quality while keeping attributes aligned to the reconstructed geometry.","pith_inferences":["If the dynamic threshold $\\tau_0$ and the model dictionary index are not counted in the bitrate, the reported rate-distortion gains are optimistic; a fair comparison should cost these side parameters explicitly.","Because LeAFNet's activation is a generic replacement for fixed activations, the same quantization-and-entropy-coding pipeline could be applied to other INR tasks such as image or video compression.","The unimodality of D1 PSNR as a function of the threshold is verified on only three trained models; testing the golden-section search on sparser and more varied LiDAR scenes would show whether the dynamic threshold remains reliable.","Using the same codec on larger-scale block-wise scenes would test whether the per-sequence optimization cost, which is not included in the reported bitrate, remains acceptable in practice."],"forward_implications":["Geometry bitstreams require roughly 44-68% fewer bits than the three MPEG anchors at matched D1 PSNR across the tested sequences.","Because decoding is just coordinate queries through the network, a single bitstream can be evaluated at arbitrary sampling resolution without re-encoding.","Improvements in neural network quantization and entropy coding transfer directly into point cloud compression gains, since the signal is only a set of weights.","Splitting geometry and attribute compression means attribute quality is bounded by the reconstructed geometry, but color optimization can proceed independently once geometry is fixed."],"supporting_citations":[{"why":"Supplies the Kolmogorov-Arnold design that LeAFNet's learnable activation layers are built on.","marker":"[21]"},{"why":"Provides the DeepCABAC arithmetic coder used to entropy-code the quantized network parameters.","marker":"[44]"},{"why":"Defines the G-PCC and V-PCC MPEG anchors that PICO is compared against.","marker":"[12]"},{"why":"Introduces the NeRF-style positional encoding that LeAFNet applies to input coordinates.","marker":"[24]"},{"why":"Gives the focal loss that the geometry occupancy classification loss is modeled on.","marker":"[20]"},{"why":"Defines the PCQM metric used for joint geometry-attribute quality evaluation.","marker":"[23]"},{"why":"Defines the BD-PSNR and BD-BR metrics used for all rate-distortion comparisons.","marker":"[4]"},{"why":"Motivates replacing B-spline bases with radial basis functions in the learnable activation.","marker":"[18]"}],"fun_headline_variants":["PICO: 4.92 dB better than MPEG via learnable latent activations","Learnable latent activations give INRs a 4.92 dB point cloud edge over MPEG","PICO: neural parameter compression beats MPEG by 4.92 dB","PICO: learnable activation INRs compress point clouds 4.92 dB better"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported bitrate counts only the quantized and entropy-coded network parameters, and the paper does not state whether the dynamic threshold $\\tau_0$ used at reconstruction and the model dictionary selection are also transmitted, so if those side bits are omitted the claimed rate-distortion gains are overstated.","fun_headline_variants_meta":{"raw":{"variants":["PICO: 4.92 dB better than MPEG via learnable latent activations","Learnable latent activations give INRs a 4.92 dB point cloud edge over MPEG","PICO: neural parameter compression beats MPEG by 4.92 dB","PICO: learnable activation INRs compress point clouds 4.92 dB better"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001141,"raw_usage":{"total_tokens":4744,"prompt_tokens":960,"completion_tokens":3784,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":3691}},"tokens_in":576,"tokens_out":3784,"duration_ms":22650,"temperature":1.0,"reasoning_tokens":3691,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:48:01.158301+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the rate-distortion curves for the four 8iVFB sequences while charging bits for the dynamic threshold (one scalar per sequence) and the model dictionary index (a few bits per rate point); if the BD-PSNR gain over G-PCC (octree) drops below the reported 7.09 dB average, the central claim fails. Alternatively, run the decoder with a fixed threshold derived only from the transmitted model and measure the resulting D1 PSNR loss.","supporting_citations":[{"cited_title":"Hou, and Max Tegmark","cited_arxiv_id":null,"evidence_quote":"Supplies the Kolmogorov-Arnold design that LeAFNet's learnable activation layers are built on."},{"cited_title":"DeepCABAC: Context-adaptive binary arithmetic coding for deep neural network compression","cited_arxiv_id":"1905.08318","evidence_quote":"Provides the DeepCABAC arithmetic coder used to entropy-code the quantized network parameters."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the PCQM metric used for joint geometry-attribute quality evaluation."}],"review_version":1}