PaperlayerView exact revision on arXiv

Source-linked AI summary

On Task Scope and Information Retention in Source Coding

Alireza Furutanpey, Kerstin Bunte

arXiv:2609.37575v1cs.ITcs.LGeess.IVmath.IT

TL;DR

The paper argues that the CfM/CfH distinction does not determine which information a codec may discard; the supported tasks and their constraints do. It extends source and feature coding to finite task families and analyzes restricted encoder observations. It shows that feature coding can match source coding for one task yet require more bits when another task is added.

  • Problem

    The CfM/CfH distinction is misleading as a basis for deciding what information a codec may discard, because receiver identity alone does not determine required source distinctions.

    Source: §1 · p. 1

  • Method

    The paper extends source and feature coding to finite task families and characterizes when restricting the encoder observation preserves the minimum rate and when it incurs a penalty.

    Source: §1 · p. 2, §7 · p. 12

  • Results

    Feature coding can match source coding for the original task yet require more bits when another task is added, despite unchanged optimal uncompressed predictions.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Compression savings apply only to specified predictions, losses, tolerances, encoder observations, and decoding constraints; preserving each task’s optimal uncompressed prediction need not preserve distinctions needed to allocate coding errors efficiently.

    Source: §7 · p. 12

    Measured rates characterize the tested codecs and fitted readouts.

    Source: §6.2 · p. 11

Abstract

from arXiv · show

We argue that dividing codec design into Coding for Machines (CfM) and Coding for Humans (CfH) is a misleading distinction for deciding what information a codec may discard. Receiver identity does not determine admissible information loss. The required rate depends on task scope, including the predictions to support, their losses and tolerated risks, the encoder observation, and the permitted decoding procedures. Notably, a machine task may have a higher minimum rate than a restricted human decision. Rate savings on selected machine tasks apply only to the stated requirements, not to an intrinsic ordering by receiver type. We extend source and feature coding to finite task families, derive when restricting the encoder observation preserves the minimum rate, and show that equality between source and split-feature coding rates can no longer hold as the task scope expands.

1 Introduction

The introduction argues that the CfM/CfH distinction is misleading for deciding what information a codec may discard: required information depends on supported task scope, not receiver identity. Its example and contributions show why codec claims must remain tied to those requirements.

1 Introduction

  • A cloud feature codec must specify its supported task scope, because savings on designer-selected tasks do not establish suitability for clients whose tasks were unknown at design time.The system model has a resource-constrained client transmitting compact representations to a shared, better-provisioned server.Source: §1 · p. 1
  • Codec design should be framed by the predictions, tolerated risks, encoder observations, and decoding constraints it supports, rather than by whether the receiver is human or machine.The paper notes that a restricted human decision may require less information than a machine task with broader predictive requirements.Source: §1 · p. 1
  • A machine task can require more bits than a restricted human decision when its lower prediction risk depends on distinctions people cannot reliably perceive.A reconstruction for general viewing may need broad perceptual fidelity, while expert inspection may need only decision-relevant evidence to remain interpretable.Source: §1 · pp. 1–2
  • At block 16, species-pair accuracy is 93.3% from native features and 82.5% from reconstructed RGB, while ImageNet decisions remain unchanged.The figure shows Indigo Bunting and Blue Grosbeak inversions across blocks; the chestnut patch becomes less visible at later blocks.Source: §1 · p. 2, Figure 1 · p. 2
  • The paper extends source and feature coding to finite task families, characterizes when restricted encoder observations preserve the optimum, and demonstrates rate penalties when tasks expand.Learned finite codes demonstrate the penalty in completely transmitted messages.Source: §1 · p. 2

2 Task scope and compression requirements

This section makes rate comparisons conditional on what the decoded message must preserve and formalizes task scope through predictions, losses, tolerances, observations, and permitted joint outputs. It positions that framework among work on task-specific fidelity, preserved decisions, coding interfaces, and reusable representations.

2.1 Task scope and rate–distortion theory

  • Compression problems with the same source and receiver can have different minimum rates when their reconstruction alphabets or distortion requirements differ.The rate–distortion formulation uses expected distortion D and mutual information measured in bits.Source: §2.1 · p. 3
  • The framework represents a finite task family with reference outputs, one risk constraint per task, and permitted tuples of joint decoded predictions from a shared message.The encoder may observe the full source X or a fixed restricted feature Y; the permitted joint prediction set need not equal the set of reference-output tuples.Source: §2.1 · p. 3
  • Restricting an observation can raise the smallest attainable prediction risk when it merges source states with different conditional target distributions.A risk requirement below the floor attainable from the observation cannot be met at any rate; a feature is sufficient for a target when it preserves the conditional target distribution.Source: §2.1 · pp. 3–4
  • Even when two predictions share a Bayes decision, different conditional error costs can affect how a code allocates errors, and multiple tasks can further restrict that allocation.The section reviews prior work on task families, preserved decisions, and coding rates, while noting that equal optimal predictions across tasks do not imply equal coding-error costs.Source: §2.1 · p. 4, §2.2 · p. 4

2.2 Related Work

  • Rate comparisons across task-oriented compression, coding interfaces, and reusable representations must specify whether they require codec reuse, an additional prediction from an existing message, or a newly fitted decoder.The related work also covers mixed fidelity objectives and methods using shared or task-specific messages.Source: §2.2 · p. 4

3 Comparing source and feature coding

This section compares source coding, split-feature coding, and direct-feature coding under common task-risk tolerances and permitted prediction tuples. It derives conditions for equality and bounds on rate penalties, showing how expanded task families and encoder objectives affect retained information.

3 Comparing source and feature coding

  • Source coding, split-feature coding, and direct-feature coding differ in whether the encoder codes X or Y and whether prediction uses a source or feature reconstruction.Their minimum rates are denoted RX, RY, and RXY, respectively.Source: §3 · p. 5

3.1 Jointly attainable predictions

  • Matching each task’s separately available outputs is insufficient if the reconstruction alphabets do not permit every required combination of joint predictions.Using a common range of permitted prediction tuples lets the three rates be computed through the same prediction optimization.Source: §3.1 · p. 5

3.2 Encoder observation and task risk

  • An encoder observing X can compute Y, so direct-feature coding equals source coding and is no worse than split-feature coding; for Harell’s fixed target, source and split-feature rates also agree.With additional targets, conditional costs can vary among inputs sharing a feature, and whether that matters depends on tolerated risks.Source: §3.2 · p. 5

3.3 Exact equality and rate penalty

  • For prescribed observations, equality holds exactly when an optimal source marginal meets the stated conditional-cost equality condition for source states merged by the feature.The condition concerns relative costs among predictions used by an optimal channel; identical Bayes predictions alone do not ensure equal rates.Source: §3.3 · p. 6
  • Under feasible risk constraints and finite dual-optimal multipliers, the paper bounds the source–feature rate penalty using differences between their objectives.An expanded task family can violate the equality criterion even when the original task and its risk tolerance remain unchanged.Source: §3.3 · p. 6
  • A source-observing encoder’s training objective alone does not determine what information its fitted message retains.Although a target-dependent channel can be constructed at the unrestricted optimum when distortion depends on X only through T, a fitted encoder need not implement it.Source: §3.3 · p. 6

4 Finite families at nonzero risk

Across finite task families, required rates depend on task risks and encoder observations, not simply the number of tasks or their uncompressed Bayes predictions. Restricting the observation can raise rates when additional tasks change how errors should be allocated, while some task requirements eliminate that advantage.

4.1 Noisy threshold tasks

  • From Y, the additional-task risk has a floor because S = 2 and S = 3 share Y = 1, giving minimum underlying-bit error 1/6.At or above this risk floor, transmitting C meets both requirements; below it, source-observing encoders additionally encode B when C = 1.Source: §4.1 · p. 7
  • 0.54085 bits per sample is the rate difference at ϵ = 0.05 and zero allowance; preserving lower errors costs less than encoding irrelevant sensor variation.The source-observation rate increment is smaller than H(X | V) = 1 + H(N).Source: §4.1 · p. 7
  • Replacing B with B′ makes both decisions functions of Y, yet their joint Bayes predictions require 1.58496 bits per sample versus 1.45915 for (C, B).The required predictions from a reduced observation therefore depend on task identity, not task count or the minimum rate from X alone.Source: §4.1 · p. 7

4.2 Retaining the original task while adding a requirement

  • The source-versus-feature rate gap is 0.18336 bits per sample, within [0.16642, 0.19660]; optimal channel original-task risks are 0.29456 and 0.15077, both meeting the unchanged requirement.With exact original-task recovery (D0 = 0), both encoders require H(Z) bits per sample and can attain the additional task’s Bayes risk of 0.45 by transmitting Z.Source: §4.2 · p. 8

4.3 Task-dependent error costs

  • For complementary tasks, averaging their risk constraints makes constant error allocation optimal, and the stated conditional-cost criterion extends to finite task families.When both risks are imposed, concentrating errors in either confidence state no longer yields the single-task advantage.Source: §4.3 · p. 8
  • 0.53100 bits per sample is the minimum source rate for both risks at 0.30, eliminating the confidence-observation advantage despite the unchanged Bayes tuple (Z, Z).For task 1 alone at D = 0.30, rates are 0.33680 bits per sample from X and 0.53100 from Y.Source: §4.3 · pp. 8–9

5 Conditional error costs in finite coding

Finite codebooks preserve the full-observation rate advantage under both requirements, including framing and padding. The advantage reflects allocating more errors to confidence states where their additional-task cost is lower; controls show that confidence matters when conditional costs vary.

5 Conditional error costs in finite coding

  • 0.81251 bits per source symbol is below the 0.81876-bit reduced-observation minimum, while all three full-observation codes satisfy both risk tolerances.The strict advantage persists for finite codebooks when framing and padding bits are counted.Source: §5 · p. 9
  • Full encoders allocate more original-task errors to negative-confidence states, where predicting class 2 incurs less additional-task loss.Reduced selectors cannot distinguish confidence states and exceed the additional-task tolerance on the same codebooks.Source: §5 · p. 9
  • When the additional target law is averaged over confidence, full and reduced selectors produce identical packets; confidence changes codes when conditional costs vary.The controls also find identical packets for the original deterministic task alone.Source: §5 · p. 9
  • Figure 4 rows list fitting losses and complete validation rates; open and filled markers show validation and selected-test means, with dashed lines marking allowances.Bars show conditional 95% paired-image intervals and small dots show individual fits.Source: Figure 4 · p. 10

6 Requirements for reusable image codes

The image studies test whether codes fitted for narrower losses can meet additional prediction requirements, and whether decoder choice changes what an unchanged message supports. Results depend on the specified risks and decoding procedures, not just the fitted loss or encoder observation.

6 Requirements for reusable image codes

  • The image-code rates are measured in bpp from complete messages, and the stated classification requirements allow at most three percentage points of coarse-accuracy loss and five points of fine-accuracy loss.The complete messages include the main latent, hyperlatent, and header; shared model parameters are stored separately.Source: §6 · p. 10

6.1 Coarse supervision and fine prediction

  • Added-fine coding meets both test accuracy-drop allowances at 0.26717 bpp, whereas the coarse-only codec exceeds both allowances at 0.19772 bpp.Both objectives selected the same codec on validation; the codecs use the same early-tensor observation, isolating the preservation objective.Source: §6.1 · p. 10

6.2 Preservation and adapted decoding

  • Across teachers, selected logit-preserving codes have smaller coarse-accuracy drops at lower complete rates than feature-preserving codes, while refitted fine accuracies are similar.Nearby lower-rate feature-preserving candidates still have higher rates and approach or exceed the coarse-drop allowance.Source: §6.2 · p. 11, Figure 5 · p. 11
  • Refitting the fine classifier improves fine accuracy without changing the logit-preserving message, showing that its inherited classifier's errors do not establish that useful fine-label information is absent.Fine distinctions can remain recoverable without fine-label supervision during codec fitting when the decoder is adapted.Source: §6.2 · p. 11, Figure 5 · p. 11

6.3 Task losses and decoder choices

  • Near 0.03 bpp, depth-semantic and depth–edge codes have similar depth risks, while semantic risk is lower for depth-semantic coding and edge risk is lower for depth–edge coding.The displayed depth–edge candidate satisfies all three moderate requirements without semantic supervision during codec fitting; across three fits, mean risks meet the requirements although only one fit individually does.Source: §6.3 · p. 12, Figure 6 · p. 12
  • Reconstructing RGB and using frozen reference predictors lowers depth and semantic risks without changing the message; with the latent edge decoder retained, all three stringent requirements are satisfied.Depth risk moves from above the moderate threshold to below the stringent threshold with the frozen predictor on reconstructed RGB.Source: §6.3 · p. 12, Figure 6 · p. 12
  • All three depth-only fits meet the loose semantic cross-entropy requirement, yet their mean mIoU is below 0.01, showing that probability-loss compliance does not establish categorical agreement.Categorical requirements must specify both the metric and the evaluated classes; the uncompressed reference's test mIoU also falls below the added floor.Source: §6.3 · p. 12

7 Conclusion

The conclusion rejects receiver identity as a guide to admissible information loss. It argues that task scope determines rate requirements, and that restricting observation may preserve the minimum rate for one task but increase it for an expanded task family.

7 Conclusion

  • The paper extends source and feature coding to finite task families, characterizes when restricting encoder observation preserves the minimum rate, and bounds the resulting penalty.Theoretical constructions and learned finite codes show that restricting observation can preserve the rate for one task but increase it for an expanded family.Source: §7 · p. 12
  • Receiver identity alone does not determine which source information a codec may discard; required information depends on the specified predictions, losses, tolerated risks, encoder observation, and decoding constraints.Compression savings apply relative to those requirements, not as an ordering of minimum rates by receiver type.Source: §7 · p. 12
  • With adapted decoding, additional risk requirements can be satisfied without changing the transmitted message or its rate.Preserving each task's optimal uncompressed prediction need not preserve distinctions needed to allocate coding errors efficiently across task risk constraints.Source: §7 · p. 12

A Coding results and proofs

The appendix relates coding rates to the predictions jointly attainable under a task family and to what the encoder observes. It gives conditions for equal rates and explains why enlarging the task family can invalidate source–feature rate equality.

A.1 Jointly attainable predictions and equality of coding rates

  • Replacing a reconstruction with its permitted prediction tuple preserves task risks and cannot increase mutual information; choosing a representative reconstruction for each tuple gives the converse construction.This lets the source, direct-feature, and split-feature rates be computed using prediction-channel optimization.Source: §A.1 · p. 15
  • For a deterministic single target depending on X only through Y, averaging the reconstruction channel over X given Y preserves distortion and does not increase mutual information, yielding equal direct and split rates.Source equality also requires an attainable-output condition or a replacement with no greater distortion; across several tasks, one replacement must serve every task.Source: §A.1 · p. 16

A.2 Prescribed observation

  • The Y-dual objective can be zero even when JY > JX, so this quantity alone does not certify equality of the rates.The stated rate bounds compare the dual expressions at the source and reduced-observation multipliers.Source: §A.2 · p. 16
  • At any tolerance feasible from Y, the source and reduced-observation rates are equal exactly when an X-optimal constrained prediction channel factors through Y.Equivalently, composing a Y-optimal channel with the observation map is X-optimal; if the tolerance is infeasible from Y, its rate is infinite.Source: §A.2 · p. 16

A.3 Three-class conditional error costs

The three-class construction exhibits a rate gap when an additional task is added, despite unchanged optimal uncompressed predictions. The appendix also gives limits on information beyond a preservation target and a prediction-only transmission rate for a restricted tuple.

A.3 Three-class conditional error costs

  • With zero–one loss, both observations have the unique Bayes prediction tuple (Z, Z), with risks (0, 0.45).Every conditional class probability is positive, and the minimum Bayes margin is 0.04875.Source: §A.3 · p. 17
  • For β > βX ≃ 0.18152990, the source-optimal marginal assigns positive probability to prediction 2 and JX < JY; for 0 < β ≤ βX, JX = JY.In the equal-rate range, the conditional cost difference for predictions 0 and 1 is independent of K; beyond it, the source-optimal marginal uses prediction 2 and the difference depends on K.Source: §A.3 · p. 17
  • At tolerances (D0, D1) = (0.35, 0.499), the family-rate gap is approximately 0.183358557763 bits per sample.The family includes all nine prediction pairs; projecting onto the additional task gives a lower bound, attained by using the same prediction for both tasks.Source: §A.3 · p. 17
  • At D0 = 0, every feasible message determines Z and needs at least H(Z) bits per sample; sending Z attains additional risk 0.45, so both family rates equal H(Z) for every feasible D1 ≥ 0.45.For the additional task alone, the interval of equal rates ends near 0.00199467 bits per sample.Source: §A.3 · p. 17

A.4 Information beyond a preservation target

  • For a complete prefix-coded message reconstructing a deterministic target at distortion D, a rate near the target rate–distortion function limits additional information beyond that target.The image-code rates are not compared with an estimated target rate–distortion function.Source: §A.4 · p. 18
  • If only the known coarse-and-fine prediction tuple is required, computing both predictions at the encoder allows transmission in 11 fixed-length bits per image before framing.Source: §A.4 · p. 18

A.5 Noisy threshold tasks

For the two noisy threshold tasks, the section characterizes their joint minimum-rate requirements and proves them using converse bounds and matching constructions. It also compares requirements defined from different reference observations.

A.5 Noisy threshold tasks

  • All three rates are infinite if either task requires risk below ε.Source: §A.5 · p. 18
  • The converse combines bounds on errors in C and B; achievability encodes C losslessly and rate-distortion codes the positions with C = 1.Source: §A.5 · p. 19
  • At the 1/6 underlying-bit error floor, transmitting C attains HC, which the coarse task also requires.Source: §A.5 · pp. 18–19
  • At ε = 0.05 and zero allowance, using X-reference risks requires 0.54085 bits per sample more than using V-reference risks.The additional bits are fewer than H(X | V), which includes independent sensor variation.Source: §A.5 · p. 19
  • The full two-task rate region is obtained by solving a finite-alphabet rate–distortion problem with two constraints and observation-specific distortion matrices.For noisy targets, each underlying-bit distortion d is transformed to ε + (1 − 2ε)d.Source: §A.5 · p. 19

A.6 Conditional error profiles

For finite task families, the section characterizes when encoder confidence information changes the minimum rate and shows that optimal joint predictions can be restricted to a common prediction. It also reports numerical and operational coding checks of these results.

A.6 Conditional error profiles

  • For 0 < e* < 1/2, source and restricted-observation rates are equal exactly when 1 lies in the convex hull of the active tasks’ normalized error profiles; otherwise RX < RY.The optimum is attained by a channel making the same prediction for every task, even when all joint prediction tuples are permitted.Source: §A.6 · p. 20
  • At a constant optimum, tasks with normalized allowance above e* have slack constraints and zero multipliers; equal fractional allowances make every task active.Repeating a task does not change the convex-hull condition.Source: §A.6 · p. 20
  • For three equiprobable confidence states, the cyclic family has equivalent full- and reduced-observation rates, while every nonempty task subset has reduced-observation rate 1 − h2(0.1) ≈ 0.531 bits.The uncompressed Bayes predictions remain Z.Source: §A.6 · p. 22

B.1 Finite-model numerical comparisons

  • Across 16 finite-model cases, the full-product and common-prediction minima differ by at most 4.84 × 10−8 bits.The cases also test the whole-family and active-family convex-hull conditions.Source: §B.1 · p. 22

B.2 Three-class operational coding

  • The three-class operational assessment reports largest full-observation upper risk bounds of 0.25248 for the original task and 0.49846 for the additional task.The assessment uses simultaneous 99% coverage for 22 component means.Source: §B.2 · p. 23

B.3.1 Source laws, fitting, and messages

  • The source laws use independent fair decision and confidence bits, and the fitting procedures compare known-loss, fitted-loss, and reduced-observation encodings.Candidate assignments are scheduled independently of the source and transmitted with an eight-byte header and packed indices.Source: §B.3.1 · p. 23

B.3.2 Assessment and score precision

  • Assessment intervals have approximate simultaneous 95% coverage conditional on fitting and selection, with a rate-equivalence margin of 1/64 bit per decision.Fitted-score differences near 10−7 can change assignments at exact known-cost ties, motivating score rounding.Source: §B.3.2 · p. 23

B.3.3 Finite-code sensitivity and fixed-codebook comparisons

Finite-code comparisons show that single-task savings persist at both tested lengths, while increasing task-family scope can raise the full-observation rate. Fixed-codebook comparisons also quantify the benefit of observing confidence.

B.3.3 Finite-code sensitivity and fixed-codebook comparisons

  • The single-task saving persists at lengths 12 and 16; single-task, duplicate-task, and loose-pair rates overlap, while confidence corruption reduces the saving until it vanishes at τ = 1/2.Tight-pair rates remain equivalent at all three corruption levels.Source: §B.3.3 · pp. 23–24, Figure 7 · p. 24
  • As three-state profiles are added, the full-observation rate increases; for the cyclic family, full- and reduced-observation rates are equivalent and uncompressed predictions remain unchanged.Source: §B.3.3 · p. 24
  • With eight-bit indices, observing confidence reduces single-task risk by 0.03779–0.03802 on fixed length-16 codebooks.With ten-bit indices, full-observation risks span 0.272730–0.272888, compared with 0.308054–0.308321 without confidence.Source: §B.3.3 · p. 24

B.3.4 Additional-task risk forecasts

The section forecasts additional-task risk from conditional error probabilities for an already fitted single-task code. Forecasts closely match independent assessments and show when stricter requirements need another code.

B.3.4 Additional-task risk forecasts

  • For a fixed binary code, conditional error probabilities and task-specific conditional costs determine another task’s risk.Source: §B.3.4 · p. 24
  • Across 252 forecasts, the largest absolute component-risk error on independent assessment is 0.001128, and forecast and assessment classifications agree under both stated rules.Source: §B.3.4 · p. 24
  • Length-16 codes have complementary-task risk about 0.420, below the loose tolerance and above the tight tolerance; stricter requirements can therefore require another code despite unchanged uncompressed Bayes predictions.Source: §B.3.4 · p. 24

B.4 ImageNet task requirements

The section tests whether codecs meet multiple task-risk requirements across ImageNet, CIFAR-100, and Taskonomy, then examines input-independent mixtures and feature reconstruction. Results show that adding tasks can change which codes satisfy requirements, and that selected outcomes depend on the evaluation setting.

B.4 ImageNet task requirements

  • For the selected added-fine code, simultaneous upper accuracy-drop limits were 2.575 coarse points and 4.367 fine points, both below their respective three- and five-point allowances.Source: §B.4 · p. 25

B.5.2 Preservation selection and test assessment

  • The selected feature-preserving group for Teacher C and lower-rate candidates for teachers A and C failed joint assessment, while other assessed mean accuracy drops met allowances.Their coarse-drop intervals were [2.310, 3.135], [2.470, 3.365], and [2.805, 3.705] points.Source: §B.5.2 · p. 26
  • 26.35%, 36.38%, and 25.73% rate reductions for teachers A, B, and C accompanied selected logit-minus-feature fine-accuracy intervals within the one-point equivalence margin.For teachers A, B, and C, the intervals were [−0.640, 0.280], [−0.835, 0.055], and [0.050, 0.928] points.Source: §B.5.2 · p. 26

B.5.3 Observation and initialization controls

  • The earlier Teacher A comparison found randomly initialized early-observation and logit-observation codecs achieved 61.43% and 41.22% test fine accuracy at 2.770 and 0.724 bpp, respectively, but differing analysis architectures prevent isolating observation cost.Another randomly initialized logit-loss fit had 66.00% validation coarse accuracy versus the 85.68% reference. Fine-supervised controls had lower coarse accuracy than focal settings, precluding an advantage claim at matched coarse risk.Source: §B.5.3 · p. 27

B.6.2 Candidate selection and fit variation

  • No candidate met stringent Taskonomy test requirements including semantics, and both selected codes and the uncompressed reference fell below the moderate object-mIoU floor.The hardest building exceeded the moderate depth-loss tolerance for every fit; doubling newly fitted decoder width from 48 to 96 left primary feasibility decisions unchanged. Retrospectively restricting test mIoU to validation-present classes raised the reference and mean DS, 3 results above the moderate floor.Source: §B.6.2 · p. 28, §B.6.4 · p. 30

B.6.3 Input-independent mixtures

  • Five of twelve mixtures and eleven of twelve single-code selections met test requirements; all six mixtures with an edge requirement exceeded its tolerance.The three DS mixtures and depth-only choices at moderate and loose tolerances met their requirements; the moderate DS mixture used frequencies approximately 0.346 and 0.654 and had test depth and semantic risks (0.080843, 0.793438).Source: §B.6.3 · p. 29

B.6.5 Complete message lengths and decoded examples

  • Taskonomy rates count entropy-coded latents and a 40-byte header, excluding model storage and decoder computation.Decoded examples compare the base D, 30, DS, 3, and DE, 3 codes on the first validation image from each building.Source: §B.6.5 · p. 30

Related papers

Loading 2609.37575v1…