Pith. sign in

REVIEW 3 major objections 5 minor 35 references

FiFTy: Large-scale File Fragment Type Identification using Neural Networks

T0 review · 3 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A compact 1-D CNN with a trainable byte-embedding layer identifies file types from raw blocks, reaching 77.5% accuracy on 75 classes and running an order of magnitude faster than the previous tool.

desk verdict Useful empirical contribution, but the headline accuracy may be inflated by a non-file-disjoint split; worth citing for the dataset and architecture, and worth sending to peer review with a request for a file-level re-run. read the letter →

arxiv 1908.06148 v2 pith:3FQJIP25 submitted 2019-08-16 cs.CR cs.MM

Govind Mittal , Pawel Korus , Nasir Memon This is my paper · ORCID
classification cs.CRcs.MM
keywords filefragmenttypeidentificationmemoryforensicsdatacarvingconvolutionalneuralnetworkbyteembeddingfile-typeclassificationFFT-75dataset
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that a compact one-dimensional convolutional network reading raw bytes can identify the file type of a memory fragment faster and more accurately than classifiers built on hand-crafted statistical features. On a new balanced benchmark with 75 file types, the proposed FiFTy model reaches 77.5% average accuracy on 4,096-byte blocks while processing about 38 seconds per gigabyte, compared with 69% accuracy at about 9 minutes per gigabyte for the leading SVM-based tool under the same test conditions. The authors' conclusion is that end-to-end learned byte statistics replace explicit feature extraction in file-type classification, and that the accompanying FFT-75 dataset is the largest and most diverse public benchmark for this task.

What carries the argument

The load-bearing component is the embedding layer: a learned lookup table that maps each of the 256 byte values to a dense real-valued vector before a stack of one-dimensional convolution and max-pooling blocks processes the sequence. This replaces sparse one-hot encoding and lets the network compress the input while convolutional filters learn characteristic local byte transitions. The final representation comes from average pooling with dropout and two fully connected layers, and per-scenario hyper-parameters are selected by the Tree-structured Parzen Estimator.

What would settle it

Recompute FFT-75 accuracy with a file-level split in which every block of a given source file is assigned wholly to training, validation, or testing, and compare the resulting FiFTy accuracy against the reported 77.5%; a substantial drop would show that the block-level shuffle inflated the headline generalization numbers.

Watch

Extended reading notes

Core claim

The central discovery is that a trainable embedding layer—one small real-valued vector per byte value, used in place of a 256-entry one-hot code—allows a compact one-dimensional convolutional network to learn the byte-transition patterns that distinguish 75 file types directly from raw blocks. On 4,096-byte blocks the model averages 77.5% accuracy across all classes, on 512-byte blocks it averages 65.6%, and in two-class photo-carving scenarios it reaches 99% to 99.6% JPEG detection accuracy. FiFTy also beats a neural network trained on byte co-occurrence matrices and one trained on global statistics, in both accuracy and runtime.

Load-bearing premise

The load-bearing premise is that randomly shuffling individual byte blocks into training and test subsets creates independent samples; blocks from the same original file can appear on both sides, so the network could memorize file-specific patterns and the 77.5% and 65.6% accuracies would overstate performance on files never seen during training.

Editorial extensions

If this is right

  • File carvers and memory-forensics tools can drop explicit feature extraction and still classify fragments, removing the largest runtime bottleneck of older systems.
  • At roughly 38 seconds per gigabyte on a server GPU, whole-disk triage of memory dumps becomes practical for high-volume investigations.
  • FFT-75 gives the community a balanced, publicly available 75-type benchmark with both 512-byte and 4,096-byte blocks, making future tools directly comparable.
  • The focused scenarios show that a two-class JPEG-versus-other detector can exceed 99% accuracy, which is directly relevant to photo carving on memory cards from cameras and IoT devices.
  • The remaining confusions are concentrated in predictable groups—HEIC/MOV, high-entropy archives, and compound office formats—so downstream carving logic can be designed around them.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The reported accuracies may be optimistic because the FFT-75 split shuffles individual blocks rather than whole files, so blocks from one source file can appear in both training and test sets; a stricter file-level split is the natural follow-up test.
  • If the result transfers to genuinely unseen files, the embedding-plus-convolution recipe should also work for other raw-byte classification tasks, such as network protocol identification or malware-family detection.
  • A natural extension is to combine FiFTy's per-block probabilities with context across neighbouring blocks, for example a conditional random field or sliding-window voting, which could further reduce errors in compound formats.
  • Because the hyper-parameter analysis found that convolution kernel size and stride matter more than embedding size, smaller and faster variants could probably shrink the embedding dimension with only a small accuracy cost.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper presents FiFTy, a 1-D convolutional neural network for file fragment type identification that operates directly on raw byte blocks with a trainable embedding layer, avoiding hand-crafted features. The authors introduce FFT-75, a balanced public dataset of 75 file types, and evaluate FiFTy on six application scenarios at block sizes of 512 and 4,096 bytes. They report 77.5% average accuracy on the 75-class scenario with 4,096-byte blocks at roughly 38 sec/GB, outperforming Sceadan (69%, 9 min/GB) and two neural baselines (NN-CO, NN-GF) under the same split and hardware. Additional experiments on the external GovDocs corpus and a DFRWS carving challenge are used to probe generalization. The paper claims state-of-the-art accuracy and speed, and releases the code and dataset publicly.

Significance. If the reported results hold, FiFTy would be the best open-source file fragment type classifier to date, and FFT-75 would be the largest balanced public benchmark for the task. The paper has notable strengths: the methodology is described in detail, the comparison to baselines is performed on the same splits and hardware, the implementation and dataset are publicly available, and the external GovDocs and DFRWS evaluations provide independent evidence for the multimedia carving scenarios. The architecture is compact (around 450k parameters for the main 75-class model) and the speed advantage over feature-based baselines is substantial and plausibly robust. The main risk is the block-level train/test split, which can leak file identity and inflate the headline accuracy on unseen files.

major comments (3)
  1. [Section IV-A] The train/validation/test split is formed by shuffling individual byte blocks after sampling them from source files, rather than by partitioning at the file level. Since blocks from the same source file can therefore appear in both training and testing, adjacent blocks sharing container headers, compression state, embedded objects, or content allow the network to memorize file-specific cues. This is load-bearing because the headline 77.5% accuracy and the comparison against Sceadan and NN-CO in Table I are all computed on this split. The authors should either provide a file-disjoint split (all blocks of a file in one split) or quantify the cross-split file overlap and show that the reported accuracies are stable under a file-disjoint split.
  2. [Section IV-E and Table VI] The external generalization results weaken the claim that FiFTy 'consistently outperforms all baselines' on unseen files. On GovDocs, scenario #1 accuracy drops from 77.5% on FFT-75 to 55.9%, and on DFRWS (512-byte blocks) to 41.4%; the paper itself notes that external results are comparable only for scenarios #3-#6. The 21.6-point drop on the 75-class scenario is the pattern expected if part of the FFT-75 score comes from file-memorization rather than generalizable type discrimination. The authors should add a file-disjoint evaluation on FFT-75 or another 75-class corpus and explicitly report the gap; if the drop persists, the headline accuracy should be qualified as an upper bound for block-level memorization rather than a claim of unseen-file generalization.
  3. [Section IV-C] No error bars, confidence intervals, or repeated splits are reported for any of the accuracy numbers, despite the fact that the test set is a random sample and the effective sample size is reduced by file-level correlation. Given that the main claim rests on a single split and a small accuracy margin over NN-CO (77.5% vs 75.3%), the authors should report results over multiple random splits or provide bootstrapped confidence intervals so that the uncertainty from the sampling procedure is visible.
minor comments (5)
  1. [Section I] In the first paragraph, 'Koglomorov complexity' should be 'Kolmogorov complexity'.
  2. [Supplementary Materials] The phrase 'inlcuding source code' contains a typo; it should be 'including source code'.
  3. [Table I] The table footnotes state that training time was computed on a Tesla P40, while Section III-B states that all experiments were run on a Tesla V100; please clarify which GPU was used for training versus inference.
  4. [Table VII] The 'Total Wins' row breaks ties by runtime speed; this rule should be stated in the table caption for clarity.
  5. [Figure 5] The figure caption lists subfigures (a)-(d) for 512-byte blocks and (e)-(h) for 4096-byte blocks, but the text references the subfigures without the block size in some places; adding the block size to each subfigure label would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: FiFTy's accuracy is a measured test-set statistic, and the approach is independently re-evaluated on GovDocs and DFRWS corpora.

full rationale

The paper does not derive any predicted quantity from a fitted parameter or from a self-citation chain. The central claim (77.5% average accuracy on 75 file types with 4,096-byte blocks) is an empirical measurement on a hold-out test set: 'The blocks are shuffled to evenly distribute the file-types within the training (80%), validation (10%) and hold-out testing (10%) subsets' (Section IV-A). The baselines (Sceadan, NN-GF, NN-CO) are evaluated on the same dataset and the same split, so the relative comparison is not constructed from FiFTy's own outputs. The paper additionally tests generalization on external corpora: 'We tested generalization capabilities of FiFTy on the GovDocs corpus and a file-system memory dump from a DFRWS1 carving challenge' (Section IV-E), with results reported in Table VI. These external evaluations are independent of the FFT-75 training data and provide a check that does not reduce to the training inputs. The only notable methodological weakness is that the random block shuffle can place blocks from the same source file into both training and testing, which may inflate absolute accuracy; however, this is a generalizability/validity concern, not circularity, because the reported accuracy is measured rather than derived from the training objective or from a fitted parameter renamed as a prediction. No load-bearing self-citation is used: the FFT-75 dataset citation [21] is a public, externally usable resource, and the cited prior works are standard baselines rather than forced premises. Accordingly, no circular step meets the evidentiary bar of exhibiting an equation that reduces to its own input, so the circularity score is 0.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard supervised learning practice plus the domain assumption that byte patterns identify file types. The weakest point is the assumption that block-level shuffling creates independent train/test samples, which is likely violated by the dataset construction.

free parameters (7)
  • Embedding size = 16, 32, 48, or 64 depending on scenario (Table II)
    Selected by TPE during hyperparameter search; affects model capacity and the reported accuracy.
  • Convolution kernel size = 11, 19, 27, or 35 depending on scenario (Table II)
    Selected by TPE; the paper notes kernel size and stride had the strongest effect on validation accuracy.
  • Convolution stride = 3, 11, 19, 27, or 35 depending on scenario (Table II)
    Selected by TPE; part of the hyperparameter space that most consistently affected accuracy.
  • Number of convolutional blocks = 1, 2, or 3 depending on scenario (Table II)
    Selected by TPE; the paper observes that one block is usually not enough.
  • Max pooling size = 2, 4, 6, or 8 depending on scenario (Table II)
    Selected by TPE; stronger pooling helped for 4,096-byte blocks.
  • Number of dense units = 16, 32, 64, 128, or 256 depending on scenario (Table II)
    Selected by TPE for the hidden fully connected layer.
  • Dropout probability = 0.1
    Chosen in a prior hand-tuned experiment; not optimized by TPE.
assumptions (4)
  • domain assumption Byte-level statistics contain enough signal to distinguish 75 file types.
    Core premise of learning f from raw bytes; supported empirically by the reported results but not guaranteed.
  • domain assumption Shuffled block-level splitting yields independent training and test examples.
    Invoked in Section IV-A; likely violated because blocks from the same source file can appear in both splits, inflating measured accuracy.
  • standard math Backpropagation and TPE optimize the model as intended.
    Standard machine learning machinery used throughout; not proven in the paper but widely established.
  • domain assumption FFT-75 is representative of real carving workloads.
    The dataset is self-constructed; external GovDocs and DFRWS results show lower accuracy, so representativeness is only partial.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FiFTy: Large-scale File Fragment Type Identification using Neural Networks." pith.science (2026). https://pith.science/paper/3FQJIP25

@misc{pith2026190806148,
  author       = {Pith},
  title        = {Pith review of: FiFTy: Large-scale File Fragment Type Identification using Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3FQJIP25}},
  note         = {Machine review of arXiv:1908.06148}
}
read the original abstract

We present FiFTy, a modern file type identification tool for memory forensics and data carving. In contrast to previous approaches based on hand-crafted features, we design a compact neural network architecture, which uses a trainable embedding space, akin to successful natural language processing models. Our approach dispenses with explicit feature extraction which is a bottleneck in legacy systems. We evaluate the proposed method on a novel dataset with 75 file types - the most diverse and balanced dataset reported to date. FiFTy consistently outperforms all baselines in terms of speed, accuracy and individual misclassification rates. We achieved an average accuracy of 77.5% with processing speed of approx 38 sec/GB, which is better and more than an order of magnitude faster than the previous state-of-the-art tool - Sceadan (69% at 9 min/GB). Our tool and the corresponding dataset are available publicly online.

Figures

Figures reproduced from arXiv: 1908.06148 by the authors.

Figure 1
Figure 1. Illustration of the proposed network architecture; the presented model [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Hyper-parameter value selection (here conv. kernel size) using TPE: [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Impact of training data size on validation accuracy and loss (scenario [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Impact of hyper-parameter variation on the validation accuracy for all scenarios with 4096-byte inputs: (top) violin plots showing distribution of [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Heat maps of confusion matrices for scenario #1 of [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Confusion matrices: on Govdocs [22] for scenario #1 and #4 (ab) [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 3
Figure 3. Figure 3: Confusion matrix for FiFTy with block size of 512 bytes. Darker color means higher value and dot (.) means a non-zero value greater than 0.2% [PITH_FULL_IMAGE:figures/full_fig_p018_3.png]
Figure 4
Figure 4. Figure 4: Confusion matrix for FiFTy with block size of 4096 bytes. Darker color means higher value and dot (.) means a non-zero value greater than 0.2% [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]
Figure 5
Figure 5. Figure 5: Confusion matrix for Sceadan with block size of 512 bytes. Darker color means higher value and dot [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]
Figure 6
Figure 6. Figure 6: Confusion matrix for Sceadan with block size of 4096 bytes. Darker color means higher value and dot [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]
Figure 7
Figure 7. Figure 7: Confusion matrix for baseline NN-GF with block size of 512 bytes. Darker color means higher value and dot [PITH_FULL_IMAGE:figures/full_fig_p022_7.png]
Figure 8
Figure 8. Figure 8: Confusion matrix for baseline NN-GF with block size of 4096 bytes. Darker color means higher value and dot [PITH_FULL_IMAGE:figures/full_fig_p023_8.png]
Figure 9
Figure 9. Figure 9: Confusion matrix for baseline NN-CO with block size of 512 bytes. Darker color means higher value and dot [PITH_FULL_IMAGE:figures/full_fig_p024_9.png]
Figure 10
Figure 10. Figure 10: Confusion matrix for baseline NN-CO with block size of 4096 bytes. Darker color means higher value and dot [PITH_FULL_IMAGE:figures/full_fig_p025_10.png]
Figure 11
Figure 11. Figure 11: Confusion matrices of the remaining five scenarios with block size of 512 bytes. Darker color means higher value and dot [PITH_FULL_IMAGE:figures/full_fig_p026_11.png]
Figure 12
Figure 12. Figure 12: Confusion matrices of the remaining five scenarios with block size of 4096 bytes. Darker color means higher value dot [PITH_FULL_IMAGE:figures/full_fig_p027_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 33 canonical work pages

  1. [1]

    Sceadan: Using Concate- nated N-Gram Vectors for Improved File and Data Type Classification,

    N. L Beebe, L. A Maddox, L Liu, and M Sun, “Sceadan: Using Concate- nated N-Gram Vectors for Improved File and Data Type Classification,” IEEE Tran. Inf. Forensics and Security , vol. 8, no. 9, pp. 1519–1530, 2013

    work page 2013

  2. [2]

    Using NLP techniques for file fragment classification,

    S Fitzgerald, G Mathews, C Morris, and O Zhulyn, “Using NLP techniques for file fragment classification,” Digital Investigation , vol. 9, pp. S44–S49, aug 2012

    work page 2012

  3. [3]

    Statistical Disk Cluster Classification for File Carving,

    C. J Veenman, “Statistical Disk Cluster Classification for File Carving,” in IEEE Int. Symposium on Information Assurance and Security , 2007

    work page 2007

  4. [4]

    Sparse Coding for N-Gram Feature Extraction and Training for File Fragment Classification,

    F Wang, T.-T Quach, J Wheeler, J. B Aimone, and C. D James, “Sparse Coding for N-Gram Feature Extraction and Training for File Fragment Classification,” IEEE Tran. Inf. Forensics and Security , vol. 13, no. 10, pp. 2553–2562, oct 2018

    work page 2018

  5. [5]

    Cocost: a computational cost efficient classifier,

    L Li, U Topkara, B Coskun, and N Memon, “Cocost: a computational cost efficient classifier,” in 2009 Ninth IEEE International Conference on Data Mining . IEEE, 2009, pp. 268–277

    work page 2009

  6. [6]

    Jpgcarve: An advanced tool for automated recovery of fragmented jpeg files,

    J De Bock and P De Smet, “Jpgcarve: An advanced tool for automated recovery of fragmented jpeg files,” IEEE Tran. Inf. Forensics and Security, vol. 11, no. 1, pp. 19–34, Jan 2016

    work page 2016

  7. [7]

    File Type Identification of Data Fragments by Their Binary Structure,

    M Karresand and N Shahmehri, “File Type Identification of Data Fragments by Their Binary Structure,” in 2006 IEEE Information Assurance Workshop

    work page 2006

  8. [8]

    Predicting the types of file fragments,

    W. C Calhoun and D Coles, “Predicting the types of file fragments,” Digital Investigation, vol. 5, pp. S14–S20, sep 2008

    work page 2008

Show all 35 references
  1. [9]

    A File Fragment Classification Method Based on Grayscale Image,

    T Xu, M Xu, Y Ren, J Xu, H Zhang, and N Zheng, “A File Fragment Classification Method Based on Grayscale Image,” Journal of Computers, vol. 9, no. 8, aug 2014

    2014

  2. [10]

    JPEG Recovery Lab,

    “JPEG Recovery Lab,” https://www.disktuna.com/photo-recovery- easeus-data-recovery-vs-jpeg-recovery-lab/, Accessed Jun 20, 2019

    2019

  3. [11]

    Carving orphaned jpeg file fragments,

    E Uzun and H. T Sencar, “Carving orphaned jpeg file fragments,” IEEE Tran. Inf. Forensics and Security , vol. 10, no. 8, pp. 1549–1563, 2015

    2015

  4. [12]

    Every shred helps: Assembling evidence from orphaned jpeg fragments,

    E Durmus, P Korus, and N Memon, “Every shred helps: Assembling evidence from orphaned jpeg fragments,” IEEE Tran. Inf. Forensics and Security, vol. 14, no. 9, pp. 2372–2386, 2019

    2019

  5. [13]

    Efficient estimation of word representations in vector space,

    T Mikolov, K Chen, G Corrado, and J Dean, “Efficient estimation of word representations in vector space,” arXiv preprint arXiv:1301.3781 , 2013

    2013 arXiv

  6. [14]

    File fragment classification using neural networks with lossless representations,

    L Hiester, “File fragment classification using neural networks with lossless representations,” 2018, Bachelor Thesis, East Tennessee State University

    2018

  7. [15]

    File Fragment Classification Using Grayscale Image Conversion and Deep Learning in Digital Forensics,

    Q Chen, Q Liao, Z. L Jiang, J Fang, S Yiu, G Xi, R Li, Z Yi, X Wang, L. C Hui, D Liu, and E Zhang, “File Fragment Classification Using Grayscale Image Conversion and Deep Learning in Digital Forensics,” in IEEE Security and Privacy Workshops , 2018

    2018

  8. [16]

    Overview of the high efficiency image file format,

    M Hannuksela, E Aksu, J Lainema, and V . M Vadakital, “Overview of the high efficiency image file format,” in JCTVC-V0072. 2015

    2015

  9. [17]

    Algorithms for hyper- parameter optimization,

    J. S Bergstra, R Bardenet, Y Bengio, and B K ´egl, “Algorithms for hyper- parameter optimization,” in Advances in neural information processing systems, 2011, pp. 2546–2554

    2011

  10. [18]

    The jpeg still picture compression standard,

    G. K Wallace, “The jpeg still picture compression standard,” IEEE Tran. consumer electronics, vol. 38, no. 1, pp. xviii–xxxiv, 1992

    1992

  11. [19]

    A Fragment Classification Method Depending on Data Type,

    N Zheng, J Wang, T Wu, and M Xu, “A Fragment Classification Method Depending on Data Type,” in IEEE Int. Conf. on Computer and Information Technology Ubiquitous Computing and Communications Dependable Autonomic and Secure Computing Pervasive Intelligence and Computing, oct 2015

    2015

  12. [20]

    Data Type Classification: Hierarchical Class-to-Type Modeling,

    N Beebe, L Liu, and M Sun, “Data Type Classification: Hierarchical Class-to-Type Modeling,” in IFIP Advances in Information and Com- munication Technology, pp. 325–343. Springer Int. Publishing, 2016

    2016

  13. [21]

    File fragment type (fft) - 75 dataset,

    G Mittal, P Korus, and N Memon, “File fragment type (fft) - 75 dataset,” http://dx.doi.org/10.21227/kfxw-8084, 2019

    2019 doi

  14. [22]

    GovDocs Dataset,

    “GovDocs Dataset,” https://digitalcorpora.org/corpora/files, Accessed Jun 20, 2019

    2019

  15. [23]

    Intel Movidius Myriad X VPU,

    “Intel Movidius Myriad X VPU,” https://www.movidius.com/myriadx, Accessed Jun 20, 2019

    2019

  16. [24]

    F Chollet et al., “Keras,” https://keras.io, 2015

    2015

  17. [25]

    Making a science of model search: Hyperparameter optimization in hundreds of dimensions for vision architectures,

    J Bergstra, D Yamins, and D. D Cox, “Making a science of model search: Hyperparameter optimization in hundreds of dimensions for vision architectures,” 2013

    2013

  18. [26]

    Sequential model-based optimization for general algorithm configuration,

    F Hutter, H. H Hoos, and K Leyton-Brown, “Sequential model-based optimization for general algorithm configuration,” in Proc. Int. Conf. on Learning and Intelligent Optimization , Berlin, Heidelberg, 2011, LION’05, pp. 507–523, Springer-Verlag

    2011

  19. [27]

    Bz2 package,

    “Bz2 package,” https://docs.python.org/3/library/bz2.html, Accessed Jun 20, 2019

    2019

  20. [28]

    Zlib package,

    “Zlib package,” https://docs.python.org/3/library/zlib.html, Accessed Jun 20, 2019

    2019

  21. [29]

    Sceadan - open source implementation,

    “Sceadan - open source implementation,” https://github.com/ UTSA-cyber/sceadan

  22. [30]

    Everything you need to know about the JPEG-killing HEIF for- mat Apple is adopting,

    “Everything you need to know about the JPEG-killing HEIF for- mat Apple is adopting,” https://www.cultofmac.com/487808/heif-vs- jpeg-image-files/, Accessed Jun 20, 2019

    2019

  23. [31]

    Overview of the high efficiency video coding (hevc) standard,

    G. J Sullivan, J.-R Ohm, W.-J Han, and T Wiegand, “Overview of the high efficiency video coding (hevc) standard,” IEEE Tran. Circuits and Systems for Video Technology, vol. 22, no. 12, pp. 1649–1668, 2012

    2012

  24. [32]

    Progressive neural architecture search,

    C Liu, B Zoph, M Neumann, J Shlens, W Hua, L.-J Li, L Fei-Fei, A Yuille, J Huang, and K Murphy, “Progressive neural architecture search,” in Proc. European Conf. Computer Vision , 2018, pp. 19–34

    2018

  25. [33]

    Exploring randomly wired neural networks for image recognition,

    S Xie, A Kirillov, R Girshick, and K He, “Exploring randomly wired neural networks for image recognition,” arXiv preprint arXiv:1904.01569, 2019

    1904 arXiv

  26. [34]

    Benchmarking state-of-the-art deep learning software tools,

    S Shi, Q Wang, P Xu, and X Chu, “Benchmarking state-of-the-art deep learning software tools,” in Int. Conf. Cloud Computing and Big Data (CCBD). IEEE, 2016, pp. 99–104

    2016

  27. [35]

    Conditional random fields: Probabilistic models for segmenting and labeling sequence data,

    J Lafferty, A McCallum, and F. C Pereira, “Conditional random fields: Probabilistic models for segmenting and labeling sequence data,” 2001. IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY 13 TABLE VII FILE -TYPE BREAKDOWN OF CLASSIFICATION ACCURACY FOR ALL MODELS ; MIS...

    2007 arXiv

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.