{"id":"17d7f1ee-f305-48a5-82e3-e83f2e12aaa6","arxiv_id":"2505.23672","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A position-dependent weighted combination of filtered and unfiltered HEVC intra predictions yields a 2.0% average bit-rate reduction for All Intra coding.","lead":"PDPC blends two HEVC prediction outputs with position-dependent weights to cut intra-frame bit rate by about 2 percent. The paper explains the gains by comparing trained optimal predictors to the new simple formula, but leaves out training data and parameter values.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (8) is not a simplification of Eq. (4): it silently drops t[x,y](p_s-p_r), so the predictor actually evaluated in Tables 1–3 is ambiguous.","rationale":"I read the paper as making two connected claims: (i) PDPC can approximate the optimal linear prediction matrices H_m with a compact position-dependent formula, and (ii) this formula yields about 2% BD-rate gains over HEVC under common test conditions. The reader's concern about missing training data and parameter values is legitimate, but the more immediately checkable weakness is that the equations themselves do not define a single predictor. Eq. (8) is presented as a simplification of Eq. (4) for linear predictors, yet simple algebra shows the two differ by t(x,y)(p_s - p_r), which is nonzero in the intended setting of filtered references. The paper gives no side condition that makes the difference vanish, no erratum, and no statement that the experimental code uses a different formula. This means the central claim is not fully specified in the manuscript: a reader cannot tell which predictor produced Table 1. The empirical gains may well be real—PDPC was later adopted in the VVC standardization effort—so I am not recommending rejection. But the manuscript needs a major revision to correct or reconcile Eqs. (4) and (8), disclose the implemented formula, and ideally report the trained parameters and per-sequence results. My verdict remains CONDITIONAL, matching the reader's, hence UNCHANGED; I partially agree with the reader because both concerns reduce to reproducibility, but the equation inconsistency is concrete and internal rather than a matter of missing training-set statistics.","tokens_in":5904,"tokens_out":7916,"duration_ms":84352,"concrete_test":"Analytic check: fix N=8, take the horizontal angular predictor P(r)[x,y] = r[-1,y] and s = a·r + (1-a)·(h_2 * r) with h_2 a binomial filter and a = 0.5. For a random reference vector r, compute Eq. (4) and Eq. (8) numerically. The max absolute difference equals max t[x,y]·|p_s - p_r| and will be nonzero, proving the simplification fails. To settle the paper's claim, the authors should state which equation was encoded in HM-16.6, re-run the Table 1 All-Intra experiment with both equations on Class B, and, if Eq. (8) is the implemented form, derive it independently or set t = 0 in Eq. (4) and update Fig. 3. Disclosing the trained alpha values would also close the reproducibility gap identified by the reader.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Load-bearing concern: the derivation of the implemented predictor is internally inconsistent. Eq. (4) defines p = A + B + t·p_r + b·p_s, with b = 1 - w - t, w = (c1v-c2v)/2^(y/dv) + (c1h-c2h)/2^(x/dh), and t = (N - min(x,y))/N. Eq. (8) then claims that for linear HEVC predictors one can avoid computing both p_r and p_s and sets p = A + B + b'·p_s with b' = 1 - w. Since b' = b + t, subtracting Eq. (4) from Eq. (8) gives t·(p_s - p_r). For any nontrivial filter s = a·r + (1-a)·(h_k * r) with a != 1, p_s = P(s) = a·P(r) + (1-a)·P(h_k * r) differs from p_r = P(r), and t > 0 throughout the block (t = 1 at the top-left corner and ~1/N at the bottom-right corner). Therefore Eq. (8) is not algebraically equivalent to Eq. (4): the latter contains t·(p_s - p_r), while the former replaces it with zero. The paper never states that p_r = p_s (which would defeat the purpose of filtered references) nor that t = 0. Consequently the exact prediction formula in the HM-16.6 experiments is ambiguous; if Eq. (8) was implemented, the theoretical motivation through Eq. (4)/(11) and Fig. 3 does not apply to it, and if Eq. (4) was implemented, Eq. (8) and the claimed avoidance of computing both predictors are wrong. Either way, the 2.0% BD-rate claim in Table 1 cannot be checked against the equations in Section 3.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Position-Dependent Prediction Combination (PDPC) for HEVC intra prediction. The method computes each predicted pixel from a weighted sum of boundary reference terms and the HEVC prediction obtained from smoothed references, with weights that depend on pixel position, prediction mode, and block size. The PDPC parameters are trained by minimizing residual mean squared error on an unspecified training set. Experiments in HM 16.6 under the All Intra common test conditions report an average luma BD-rate reduction of 2.01% (Table 1), about 4% on three UHD sequences (Table 2), and 1.45% for a reduced two-parameter-set version (Table 3).","tokens_in":6335,"tokens_out":5229,"duration_ms":49345,"significance":"If the reported gains are reproducible, PDPC is a practically relevant, low-complexity, parallel-friendly extension to HEVC intra coding, and the paper's use of visualizations of optimal predictor matrices to motivate a parametric predictor is a useful methodological contribution. The authors integrate the method into the HM-KTA software, which is a concrete step toward reproducibility. However, the central claim is not yet fully supported: the derivation of the implemented predictor is internally inconsistent, and the training data, trained parameter values, and per-sequence results are omitted, so the reported average cannot be independently checked or separated from possible training/test overlap.","major_comments":[{"comment":"Eq. (8) is not algebraically equivalent to Eq. (4). Substituting b' = b + t and comparing the two expressions gives p_(8) = p_(4) + t[x,y](p_s^(HEVC) - p_r^(HEVC)), so the two formulas differ unless p_s^(HEVC) equals p_r^(HEVC). The paper does not state that the smoothed-reference prediction equals the unfiltered-reference prediction; indeed, with s defined in Eq. (10) and a general parameter a, p_s and p_r are generally different, and t[x,y] is positive throughout the block. The sentence in Section 3 claiming that 'when we employ only linear predictors... we can avoid having to compute both p_r and p_s' does not resolve the discrepancy, because linearity alone does not make the two predictor values coincide. The reader therefore cannot tell which formula was implemented in the HM-16.6 experiments: if Eq. (8) was implemented, the optimality argument around Eqs. (4) and (11) and the visualization in Fig. 3 do not apply to it; if Eq. (4) was implemented, the simplification in Eq. (8) is wrong. This issue is load-bearing because Tables 1-3 are presented as evidence for the proposed method, and the exact prediction formula needs to be unambiguous.","section":"Section 3, Eq. (11) and Section 4, Tables 1-3"},{"comment":"The manuscript does not disclose the training data, the trained parameter vector alpha, or per-sequence BD-rate results. The optimization in Eq. (11) is described only abstractly, with no information about the number or resolution of training sequences, the QP range, or whether the test sequences under common test conditions overlap with the training set. Since the parameters are explicitly fitted to minimize prediction error on training data, the reported 2.0% average could, for all the reader knows, partly reflect overfitting to the training distribution. Without these details, the central empirical claim cannot be reproduced or independently assessed.","section":"Section 4, Tables 1-3"}],"minor_comments":[{"comment":"The word 'develped' in the second bullet should be 'developed'.","section":"Section 3, bullet list"},{"comment":"The phrase 'an shown' should be 'as shown'.","section":"Section 4, first paragraph"},{"comment":"The phrase 'the the common tests conditions' contains a duplicated article and should be 'the common test conditions'.","section":"Section 4, first paragraph"},{"comment":"The rule 'dv = dh = 1, if 16 × 16 or smaller' is slightly ambiguous because block sizes are powers of two; it would be clearer to list explicitly N = 4, 8, 16 for dv = 1 and N = 32, 64 for dv = 2.","section":"Section 3, Eq. (7)"},{"comment":"The tables report only class averages and no per-sequence numbers or error bars, so the variability across sequences is unknown; reporting per-sequence results would make the two-decimal precision in the averages meaningful.","section":"Section 4, Tables 1-3"},{"comment":"The parameter index k for the binomial filter is not specified; the paper should state the set of allowed k values used in training and testing.","section":"Section 3, Eq. (10)"}],"recommendation":"major_revision","confidential_remarks":"The paper appears to describe a standards-oriented contribution, and the methodology may be sound in its original context. The main concern is the unresolved discrepancy between Eq. (4) and Eq. (8), which makes the exact tested predictor ambiguous. The missing training details and parameter values are also nonstandard for a self-contained paper; they may be acceptable for a conference contribution, but for a journal referee report they are needed to substantiate the claimed gains."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is the paper that introduced PDPC, and the 2% All-Intra BD-rate gain is probably real—later codec history backs it up. But the writeup has a real internal inconsistency between Eq. (4) and Eq. (8), and the experimental section omits the training details needed to fully trust the number. Worth refereeing, but it needs a corrected derivation and more complete reporting before I'd call it solid.\n\nWhat's new: PDPC combines HEVC's unfiltered and filtered reference predictions with position-dependent weights, using a compact parameterization (c1/c2, d, k, a) and a closed-form training objective in Eq. (11). The visualization of learned predictor matrices in Fig. 2 is a nice piece of analysis, and the parallel-computability motivation is legitimate—recursive filters are awkward on modern hardware. The measured 2% average gain over HM 16.6 under HEVC common test conditions, with larger gains on UHD, is a meaningful, directly usable result for codec work. The paper also honestly signals that later HM-KTA includes a slightly more optimized version.\n\nThe soft spots are real. First, the Eq. (4) to Eq. (8) step. Eq. (4) has t[x,y]p_r + b[x,y]p_s with t = (N - min(x,y))/N. Eq. (8) drops the p_r term and uses b' = 1 - w, which is b + t. The difference between the two expressions is t(p_r - p_s). For the filtered reference s = a r + (1-a)(h_k * r) with a != 1, p_s and p_r are not equal, and t > 0 through the block. So Eq. (8) is not algebraically equivalent to Eq. (4). The paper never states p_r = p_s or t = 0. If Eq. (8) was what went into HM 16.6, the theoretical motivation through Eq. (11) and Fig. 3 does not apply to the tested predictor; if Eq. (4) was implemented, Eq. (8) and the claimed complexity saving are wrong. Either way, the reader cannot map the reported BD-rate numbers back to the equations in Section 3. That is a load-bearing inconsistency, not a cosmetic typo.\n\nSecond, the training details. Eq. (11) optimizes alpha on an unspecified training set, and none of the trained parameter values are reported. We don't know whether the test sequences overlap the training data, whether the gains are stable across sequences, or how much variance there is. Tables 1-3 give class averages only. Those are standard omissions for a short conference paper, but they matter because the whole claim is an empirical gain. The 2% figure is plausible—PDPC later became a VVC tool—but this manuscript alone doesn't fully substantiate it. The citation pattern is otherwise fine; the recursive-filter prior work is relevant and the HM-KTA software is acknowledged.\n\nWho's this for: codec researchers and standards folks. It's a coding-tool paper, not a theory paper. I'd send it to review despite the issues, because the idea is important and the flaws are fixable. But it should not be accepted as-is; the authors need to correct the derivation and disclose training data and parameters. My recommendation: engage with it, but require a revision with the math fixed and the experimental reporting completed.","headline":"PDPC is a real and influential coding tool, but this paper's own equations don't match the tested predictor and the training details are missing.","tokens_in":6842,"tokens_out":4793,"would_cite":true,"duration_ms":44611,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The authors propose Position-Dependent Prediction Combination, a low-complexity extension to HEVC intra prediction that blends unfiltered and smoothed reference predictions with position-dependent weights, and report an average 2.0%…","keywords":["HEVC","intra-frame prediction","position-dependent prediction combination","reference sample smoothing","linear prediction","BD-rate","UHD video coding","parallel prediction"],"falsifier":"Train the PDPC parameters on one subset of the HEVC Class A-E sequences and evaluate on the held-out subset; if the held-out luma BD-rate gain drops substantially below 2%, the original result is overfit. A second check is to replace the trained $\\alpha$ with random weights of the same structure and verify that the coding gain collapses, confirming the parameters themselves carry the benefit.","tokens_in":5710,"feed_emoji":"🎥","tokens_out":7961,"duration_ms":64521,"temperature":0.7,"pith_summary":"The paper proposes an extension to HEVC intra-frame prediction called Position-Dependent Prediction Combination (PDPC). PDPC blends the standard HEVC predictor with predictions computed from a smoothed reference line, using weights that decay exponentially with distance from the top and left block borders and depend on prediction mode and block size. The claim is that this simple, non-recursive combination matches the behavior of optimal linear predictors well enough to cut bit rate by an average of 2.0% (luma) under HEVC All Intra common test conditions, with roughly 4% gains on UHD sequences. The authors argue the scheme keeps all pixel predictions computable in parallel, unlike earlier recursive-filtering approaches, while reusing HEVC's existing prediction modes.","feed_headline":"Position-dependent blending cuts HEVC intra bit rate 2%","feed_subtitle":"Weights depend on pixel position and mode, with 4% gains on UHD and fully parallel prediction.","key_machinery":"The mechanism is the PDPC prediction equation (8): $p[x,y] = \\frac{c_1^{(v)} r[x,-1] - c_2^{(v)} r[-1,-1]}{2^{y/d_v}} + \\frac{c_1^{(h)} r[-1,y] - c_2^{(h)} r[-1,-1]}{2^{x/d_h}} + b'[x,y] p_s^{(\\mathrm{HEVC})}$ with normalization $b'[x,y] = 1 - (c_1^{(v)} - c_2^{(v)})/2^{y/d_v} - (c_1^{(h)} - c_2^{(h)})/2^{x/d_h}$ and smoothed reference $s = a r + (1-a)(h_k * r)$ from Eq. (10), where $h_k$ is a binomial filter of order $k$. The parameter vector $\\alpha = \\{c_1^{(v)}, c_2^{(v)}, c_1^{(h)}, c_2^{(h)}, d_v, d_h, a, k\\}$ is selected per mode and block-size class by minimizing the trace objective in Eq. (11), which is the residual mean-squared error of the linear prediction matrix $\\tilde{H}_m(\\alpha)$ against the covariance statistics $P_m$ and $Q_m$ estimated from HEVC-encoded data. Because Eq. (8) is a linear combination with weights that depend only on pixel position, all pixels in a block can be computed independently, satisfying the parallel-computation requirement that recursive filters violate.","core_discovery":"The central discovery is that the smoothing that recursive filters applied to intra prediction can be reproduced by a position-dependent weighted sum of unfiltered and filtered reference contributions. Concretely, the paper shows that a predictor of the form of Eq. (8) — two exponentially decaying boundary terms anchored at the top and left references plus a normalized HEVC prediction term — approximates the trained optimal linear predictor matrices well enough to deliver consistent BD-rate reductions. The measured luma reduction is -2.01% on average across HEVC Classes A-E, and -4.06% on three UHD 3840x2160 sequences, relative to the HEVC reference software in All Intra main-10 configuration. The authors also demonstrate via predictor-matrix visualizations that the compact PDPC matrices reproduce the blurred directional patterns of the optimal predictors.","pith_inferences":["The paper does not disclose the training set or the learned parameter values, so an immediate reproducibility check is to retrain on a public dataset and measure the same BD-rate differences; until then the 2% average should be treated as a single-dataset estimate.","The same position-dependent blending idea could extend to later codecs with multiple reference lines or intra sub-partition tools, where the exponential weight formulas are replaced by per-mode trained weights or a small neural network.","The visualization of predictor matrices ($H_m$) as images is a transferable diagnostic: it could be used to compare any proposed low-complexity predictor against the optimal linear one before full encoder integration.","A parameter-free variant with fixed decay rates (e.g., $d_v = d_h = 1$) might trade some gain for zero signaling, which the paper's two-parameter-set results (1.45% average) suggest is already close to the four-set version."],"forward_implications":["Adopting PDPC into an HEVC-based encoder would provide roughly 2% BD-rate savings in All Intra coding with only a few extra multiply-adds per pixel and a per-coding-unit flag to select among two or four parameter sets.","The improvement grows with resolution: the three UHD sequences average -4.06% luma, suggesting the benefit is concentrated in the larger prediction blocks that dominate high-resolution content.","Because every predicted pixel is an independent weighted sum, PDPC supports full parallel pixel computation, which is difficult with the recursive filters it replaces.","PDPC does not require changes to HEVC's reference sample construction, transform, or entropy coding; it only modifies the prediction combination step."],"supporting_citations":[{"why":"Supplies the HEVC standard overview and coding structure that the proposed extension builds on.","marker":"[1]"},{"why":"Defines the HEVC intra prediction modes that PDPC reuses as its base predictor.","marker":"[2]"},{"why":"Establishes the reference-pixel configuration and smoothing filter table used in HEVC intra prediction.","marker":"[3]"},{"why":"Introduces recursive extrapolation intra prediction, the approach PDPC aims to replace with a parallel-friendly scheme.","marker":"[6]"},{"why":"Provides the binomial filters $h_k$ used in Eq. (10) to create the smoothed reference $s$.","marker":"[8]"},{"why":"Defines the HEVC common test conditions used to measure the 2.0% bit-rate reduction.","marker":"[9]"},{"why":"Supplies the BD-rate measure used to compute the reported coding gains.","marker":"[10]"}],"fun_headline_variants":["Position-dependent blending trims HEVC intra bitrate by 2%","Non-recursive predictors mimic recursive filters, cut HEVC intra 2%","Parallel intra-coding method beats HEVC: 2% bitrate, 4% on UHD","Weighted raw and filtered references shave 2% off HEVC intra rate","Fully parallel intra prediction: 2% average bitrate gain over HEVC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed 2.0% average bit-rate reduction relies on trained PDPC parameters that were optimized on an undisclosed video dataset and then evaluated on the HEVC common test sequences; if that training set is unrepresentative, the gains could shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["Position-dependent blending trims HEVC intra bitrate by 2%","Non-recursive predictors mimic recursive filters, cut HEVC intra 2%","Parallel intra-coding method beats HEVC: 2% bitrate, 4% on UHD","Weighted raw and filtered references shave 2% off HEVC intra rate","Fully parallel intra prediction: 2% average bitrate gain over HEVC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000875,"raw_usage":{"total_tokens":3727,"prompt_tokens":829,"completion_tokens":2898,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":445,"completion_tokens_details":{"reasoning_tokens":2791}},"tokens_in":445,"tokens_out":2898,"duration_ms":20093,"temperature":1.0,"reasoning_tokens":2791,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:39:06.704532+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the PDPC parameters on one subset of the HEVC Class A-E sequences and evaluate on the held-out subset; if the held-out luma BD-rate gain drops substantially below 2%, the original result is overfit. A second check is to replace the trained $\\alpha$ with random weights of the same structure and verify that the coding gain collapses, confirming the parameters themselves carry the benefit.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the HEVC standard overview and coding structure that the proposed extension builds on."},{"cited_title":"Position Dependent Prediction Combination For Intra-Frame Video Coding","cited_arxiv_id":"2505.23672","evidence_quote":"Defines the HEVC intra prediction modes that PDPC reuses as its base predictor."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the reference-pixel configuration and smoothing filter table used in HEVC intra prediction."},{"cited_title":"Intra coding of the HEVC standard,","cited_arxiv_id":null,"evidence_quote":"Introduces recursive extrapolation intra prediction, the approach PDPC aims to replace with a parallel-friendly scheme."},{"cited_title":"Video coding technology proposal by Samsung (and BBC),","cited_arxiv_id":null,"evidence_quote":"Provides the binomial filters $h_k$ used in Eq. (10) to create the smoothed reference $s$."},{"cited_title":"Improved video compression efficiency through flexible unit representation and corresponding extension of coding tools,","cited_arxiv_id":null,"evidence_quote":"Defines the HEVC common test conditions used to measure the 2.0% bit-rate reduction."},{"cited_title":"A recursive extrapola- tion approach to intra prediction in video coding,","cited_arxiv_id":null,"evidence_quote":"Supplies the BD-rate measure used to compute the reported coding gains."}],"review_version":1}