Benefits of Feature Extraction and Temporal Sequence Analysis for Video Frame Prediction: An Evaluation of Hybrid Deep Learning Models

Alberto Nogales; \'Alvaro J. Garc\'ia- Tejedor; Andrew Coney; Jose M. S\'anchez Vel\'azquez; Mingbo Cai

REVIEW 3 major objections 6 minor 16 references

Benefits of Feature Extraction and Temporal Sequence Analysis for Video Frame Prediction: An Evaluation of Hybrid Deep Learning Models

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Predicting autoencoder feature maps instead of raw frames yields 5-8x faster inference and 70-94% lower energy consumption, at a small accuracy cost.

desk verdict A systematic but flawed evaluation: the efficiency advantage rests on an unspecified baseline, and the accuracy headline is contradicted by the paper's own tables. read the letter →

arxiv 2508.00898 v1 pith:N2GXGD24 submitted 2025-07-28 cs.CV cs.AI

Jose M. Sánchez Velázquez , Mingbo Cai , Andrew Coney , Álvaro J. García- Tejedor , Alberto Nogales This is my paper · ORCID

classification cs.CVcs.AI

keywords videoframepredictionfeaturemapsautoencoderConvLSTM3D-CNNtemporalsequenceanalysishybriddeeplearningenergyefficiency

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 argues that predicting video frames at the level of autoencoder feature maps, rather than raw pixels, delivers large computational and energy savings with only a slight loss in prediction accuracy. The proposed pipeline compresses each frame with an autoencoder, predicts the next frame's feature map using a temporal model (RNN, LSTM, GRU, 3D-CNN, ConvLSTM, or RCNN), and reconstructs the frame with the autoencoder decoder. Evaluated on Moving MNIST, ICPR'04, and UCF101, the feature-map approaches require five to eight times less time per iteration and reduce energy consumption by 70 to 94% compared with a direct frame-prediction baseline, while the baseline achieves higher SSIM on all three datasets. The best temporal models are ConvLSTM and 3D-CNN, and grayscale real-world videos are the easiest to predict. The intended practical upshot is that feature-space prediction is a strong efficiency trade-off for resource-constrained video applications such as streaming, compression, and edge inference.

What carries the argument

The central object is the feature map, the low-dimensional encoding produced by the bottleneck layer of an autoencoder trained on each dataset. The pipeline has three stages: autoencoder feature extraction, temporal sequence modeling that predicts the next feature map from a window of past feature maps, and autoencoder decoder reconstruction of the output frame. The efficiency gain comes from dimensionality reduction: the temporal models operate on compact features instead of full-resolution frames. The named architectures that carry the temporal step are ConvLSTM and 3D-CNN.

What would settle it

Reproduce the comparison by training the direct baseline with the same hyperparameter grid, architecture search, and training budget as the feature-map models on the same GPU, and measure per-iteration time and power draw with a hardware meter; if the feature-map approach no longer achieves 5-8x speed and 70-94% energy savings, the central efficiency claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that feature-level prediction with autoencoder reconstruction yields large gains in computational efficiency and energy consumption while incurring slightly lower accuracy than direct frame prediction. Across the three datasets, the approach requires between five and eight times less time per iteration and cuts the total energy budget by 70 to 94%, with the largest savings on the more complex real-world videos. The autoencoder bottleneck reduces each frame to a compact representation, so the temporal models run on low-dimensional features rather than full-resolution images. In the accuracy comparison, the direct baseline scores higher SSIM on every dataset, for example 0.85 versus 0.70 on UCF101, so the contribution is an efficiency-accuracy trade-off: what the feature-map pipeline loses in perceptual fidelity it gains in speed and energy, which the authors argue is decisive for continuous-inference deployments.

Load-bearing premise

The direct-prediction baseline is a fair, comparably strong competitor, though the paper provides no architecture, loss, optimizer, hyperparameter grid, training budget, or hardware details for it.

Editorial extensions

If this is right

If the efficiency figures hold, feature-map prediction becomes a practical default for embedded and edge video systems where latency and battery life outweigh peak image quality.
The results indicate that combining an autoencoder with ConvLSTM or 3D-CNN temporal models is an effective recipe for grayscale and moderately complex video, pointing future work toward closing the accuracy gap on color real-world sequences.
A 70-94% energy reduction, if reproducible, makes feature-space prediction attractive for continuous-inference workloads such as video streaming, compression, and surveillance.
Because the direct baseline wins on SSIM for all three datasets, the paper positions the hybrid approach as a constrained-resource alternative rather than a general accuracy improvement.

Reading between the lines

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

A fairer comparison would require the direct baseline to receive the same grid-search budget and architecture tuning; without that, the reported speed and energy advantages may overstate the benefit of feature-level prediction.
The energy figures are reported as totals without specifying the measurement tool or hardware, so replicating the energy analysis with a hardware power meter on the same GPU would be a stronger test of the 70-94% claim.
Previous latent-space video prediction work already forecasts in a compressed domain, so the paper's novelty likely rests on the systematic multi-model comparison and the explicit cost-accuracy analysis rather than on the idea of feature-level prediction itself.
A testable extension would be to make the autoencoder bottleneck width adaptive per dataset or per scene, since the accuracy gap on UCF101 suggests that fixed compression may discard details needed for complex color motion.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

You should know two things about this paper. First, the abstract's accuracy claim is not supported by the paper's own numbers: it says SSIM increased from 0.69 to 0.82, but Table 7 shows the best per-dataset SSIM values are 0.7628, 0.8081, and 0.6981, and Table 10 shows a baseline that beats all of these (0.8032, 0.9023, 0.8531). Second, the central efficiency claim—5–8x speedup and 70–94% energy savings—rests entirely on a baseline that is never described. No architecture, hyperparameter grid, training budget, hardware, or energy measurement protocol is given, so those ratios are uninterpretable.

What the paper does well: it implements a clean three-stage pipeline (autoencoder feature extraction, temporal model on feature maps, decoder reconstruction) with six spatiotemporal architectures across three datasets that vary in realism and color. The grid search is systematic, and the reporting of multiple metrics (MAE, MSE, SSIM, DKL) plus the interval-based analysis of prediction quality gives a reasonably complete picture of where the method succeeds and fails. The comparison tables are honest in the sense that the raw numbers are there, and the text openly admits the baseline wins on SSIM everywhere.

Where it falls down: the baseline omission is load-bearing. Tables 11 and 12 report items like “1,267 s + 1,079 s” and “120,160 J + 292,388 J” with no explanation of what the two summands include (training? inference? feature extraction? reconstruction?). Without that, the reader cannot assess whether the speedup is a real property of latent-space prediction or an artifact of comparing against an oversized or undertuned direct predictor. The novelty claim also does not survive contact with the cited literature: latent space frame prediction with ConvLSTM already appears in Liu et al. 2021, which the paper cites. So the contribution reduces to an architecture and hyperparameter scan, which can be useful engineering guidance but is not the first demonstration of the idea.

The accuracy vs. efficiency trade-off is real and worth studying, but the paper does not provide enough methodological detail to make the efficiency half of the trade-off credible. The math and the data are not internally contradictory, and there is no sign of fabrication. The flaws are fixable: specify the baseline completely, define the time and energy components, correct the abstract, and recalibrate the novelty claim. As it stands, I cannot recommend accepting it.

I would send this to peer review because the question is legitimate and the paper has enough structure that a careful referee could pin down what is needed. But the current version should be rejected; the baseline detail alone is a sufficient reason. If the authors provide that and fix the abstract, the paper might become a modest but useful empirical contribution.

Referee Report

3 major / 6 minor

Summary. The paper proposes a three-stage pipeline for video frame prediction: an autoencoder first extracts feature maps from input frames, a spatiotemporal sequence model (RNN, LSTM, GRU, 3D-CNN, ConvLSTM, or CRNN) predicts the next feature map, and the autoencoder decoder reconstructs the predicted frame. The approach is evaluated on Moving MNIST, ICPR'04, and UCF101 after standardization to 599 sequences of 20 frames at 64x64 resolution. The authors report that hybrid models using ConvLSTM and 3D-CNN achieve the best accuracy, that the proposed method is slightly less accurate than a direct frame-prediction baseline on most metrics, and that the feature-map approach is substantially faster and more energy-efficient. The manuscript also claims that this is the first work to evaluate feature-level prediction with autoencoder-based reconstruction.

Significance. If the efficiency and accuracy trade-off were properly supported, the feature-map prediction idea could be a useful practical contribution for resource-constrained video prediction, and the systematic comparison across synthetic, grayscale, and color datasets is a reasonable evaluation structure. The paper does provide detailed hyperparameter grids, per-model MSE/SSIM tables, and a clear three-stage workflow, which are useful starting points for reproducibility. However, the central efficiency claim rests on an underspecified baseline, and the headline accuracy claim is contradicted by the paper's own tables. The novelty claim is also weakened by the paper's own citation of prior latent-space prediction work. As it stands, the contribution is not independently verifiable from the manuscript.

major comments (3)

[Abstract and Section 5] The abstract states that "SSIM metrics increasing from 0.69 to 0.82" and Section 5 reports SSIM scores of 0.72, 0.82, and 0.69, but the paper's own data do not support this. Table 7 gives best per-dataset SSIM values of 0.7628, 0.8081, and 0.6981 for the proposed models, while Table 10 gives baseline SSIM values of 0.8032, 0.9023, and 0.8531. The proposed method is therefore consistently below the baseline on SSIM for all three datasets, not improving to 0.82 as a general result. This wording must be corrected and the accuracy comparison presented without overstatement.
[Section 4.2, 'Comparison against a baseline'; Tables 9-12] The baseline is never specified: the text only says it is "a spatiotemporal Deep-Learning model trained directly on the original images." No architecture, loss function, optimizer, hyperparameter grid, number of input frames, parameter count, training budget, GPU/hardware, or energy measurement protocol is reported. Because the 5-8x per-iteration speedups and 70-94% energy savings in Tables 11-12 are ratios against this unnamed baseline, those numbers cannot be interpreted as evidence for the paper's central efficiency claim. In addition, the total execution time and energy entries concatenate two unexplained summands (e.g., "1,267 s + 1,079 s"), with no definition of which stage each term covers.
[Section 5 and Tables 11-12] The reported efficiency advantage is largely a definitional consequence of the design: the feature-map pipeline operates on a compact bottleneck representation, so fewer floating-point operations per step are expected. Without a baseline matched in parameter count, tuning budget, and accuracy, the speedup is not an empirical discovery but a restatement of dimensionality reduction. The paper should either compare against a well-tuned direct predictor with a similar total compute budget or explicitly frame the result as an expected trade-off with quantified causes.

minor comments (6)

[Section 3.3, Equations (1)-(4)] The equations and their captions are inconsistent: Equation (1) is labeled MSE but written with an absolute value and is actually the MAE formula, while the text says "In Equation 1, the metric is formalized" for MAE and later refers to Equation 2 for MSE and Equation 3 for SSIM. The equations should be renumbered and relabeled consistently.
[Section 4.2, Tables 6] There are two tables labeled Table 6: one reports MSE values for the spatiotemporal models and the other reports "Complete evaluation for the best spatiotemporal model" with only MAE values. This duplicate numbering makes it difficult to know which results correspond to which table and must be corrected.
[Section 2 and Section 3.2] The claim that "to the best of our knowledge, this is the first work to evaluate the advantages of feature-level prediction" is contradicted by the paper's own citation of B. Liu et al. (2021), which proposes latent-space frame prediction using ConvLSTM. The novelty statement should be narrowed to acknowledge this prior work.
[Section 4.2] The text mentions that K-Fold validation was applied, but no value of k or details of the cross-validation procedure are given, even though Table 6 reports mean and standard deviation values that presumably come from such a procedure.
[Table 8] The percentage statements are imprecise: 1,015 out of 1,785 frames is about 57%, not "around 50%," and the reported "range width" values do not match the interval endpoints shown (e.g., the interval [0.91, 0.83] has width 0.08, not 0.07).
[Table 1] The optimizer row reads "Adam RSMProp," which should be separated into two values for clarity, and the table would benefit from the usual punctuation separating list items.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation identified; efficiency and accuracy results are measured comparisons, not self-referential predictions.

full rationale

The paper's derivation chain is self-contained rather than circular. The three-stage pipeline (autoencoder feature extraction, spatiotemporal prediction on feature maps, decoder reconstruction) is a method definition; no stage assumes the performance or efficiency it later reports. Accuracy is evaluated on held-out test splits of three datasets against a direct-prediction baseline, and the baseline is reported to beat the proposed method on SSIM across all datasets, which is a substantive and non-circular finding. The efficiency gains in Tables 11-12 are empirical measurements of wall-clock time and energy, not fitted parameters or quantities defined by the method. Although the per-iteration speedup is an expected consequence of processing compact feature maps, the paper does not use the efficiency numbers to define or fit the approach; it reports them as observations. The only self-citation (Nogales et al., 2024) supports standard data-split and grid-search practice and is not load-bearing for the central claims. The under-specified baseline architecture is a reproducibility and comparability limitation, but it is not a circularity. Therefore no circular step is present.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper rests on standard empirical ML assumptions and tuned hyperparameters. The most fragile assumption is the fairness of the unspecified baseline, which is the sole basis for the efficiency claim.

free parameters (5)

Autoencoder latent dimensions = [64,128,256] for MNIST/UCF101; [32,64,128] for ICPR'04
Selected from grid search over [32,64,128] and [64,128,256] to minimize reconstruction MSE (Table 4).
Autoencoder loss function = L1 (MNIST), RMSE (ICPR'04), MSLE (UCF101)
Chosen per dataset via grid search (Table 4).
Best spatiotemporal model per dataset = 3D-CNN (MNIST), ConvLSTM (ICPR'04, UCF101)
Selected by validation MSE in Table 6.
Number of input frames = 5 (MNIST 3D-CNN), 3 (ICPR'/UCF101 ConvLSTM)
Selected from grid search values {3,5,10}.
Learning rate = 0.01 (MNIST), 0.0001 (ICPR'/UCF101)
Selected from {0.01,0.001,0.0001}.

assumptions (5)

domain assumption The three datasets are representative of synthetic/real and grayscale/color video prediction
Conclusions about ease of prediction are generalized from Moving MNIST, ICPR'04, and UCF101 only.
domain assumption Preprocessing, including truncation to 20 frames and resizing to 64x64, preserves the temporal and spatial structure needed for prediction
Section 4.1 asserts continuity checks but no quantitative verification is provided.
domain assumption The autoencoder bottleneck captures sufficient information for frame reconstruction and future frame prediction
The method predicts in bottleneck space; if the bottleneck discards predictive cues, downstream results degrade.
domain assumption The direct prediction baseline is a fair and comparably strong model
Baseline architecture and tuning are unspecified, so the efficiency comparison rests on an unverified parity of effort.
domain assumption Grid search over the listed hyperparameters finds near-optimal configurations
No random search or full search is described; optimality of chosen parameters is assumed.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Benefits of Feature Extraction and Temporal Sequence Analysis for Video Frame Prediction: An Evaluation of Hybrid Deep Learning Models." pith.science (2026). https://pith.science/paper/N2GXGD24

@misc{pith2026250800898,
  author       = {Pith},
  title        = {Pith review of: Benefits of Feature Extraction and Temporal Sequence Analysis for Video Frame Prediction: An Evaluation of Hybrid Deep Learning Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N2GXGD24}},
  note         = {Machine review of arXiv:2508.00898}
}

read the original abstract

In recent years, advances in Artificial Intelligence have significantly impacted computer science, particularly in the field of computer vision, enabling solutions to complex problems such as video frame prediction. Video frame prediction has critical applications in weather forecasting or autonomous systems and can provide technical improvements, such as video compression and streaming. Among Artificial Intelligence methods, Deep Learning has emerged as highly effective for solving vision-related tasks, although current frame prediction models still have room for enhancement. This paper evaluates several hybrid deep learning approaches that combine the feature extraction capabilities of autoencoders with temporal sequence modelling using Recurrent Neural Networks (RNNs), 3D Convolutional Neural Networks (3D CNNs), and related architectures. The proposed solutions were rigorously evaluated on three datasets that differ in terms of synthetic versus real-world scenarios and grayscale versus color imagery. Results demonstrate that the approaches perform well, with SSIM metrics increasing from 0.69 to 0.82, indicating that hybrid models utilizing 3DCNNs and ConvLSTMs are the most effective, and greyscale videos with real data are the easiest to predict.

Figures

Figures reproduced from arXiv: 2508.00898 by the authors.

**Figure 2.** Examples of the intervals’ limits for each dataset Comparison against a baseline. Regarding the previous results, our approach confirms that the feature-map-based approach not only delivers good performance in the latent space but also retains its good results once the images are restored. To assess its effectiveness in the prediction task, we compute a baseline: a spatiotemporal DeepLearning model trained directl… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 14 canonical work pages

[1]

Sánchez Velázquez1, Mingbo Cai2, 3, Andrew Coney1, Álvaro J

Benefits of Feature Extraction and Temporal Sequence Analysis for Video Frame Prediction: An Evaluation of Hybrid Deep Learning Models Jose M. Sánchez Velázquez1, Mingbo Cai2, 3, Andrew Coney1, Álvaro J. García-Tejedor and Alberto Nogales1 1 CEIEC Research Institute, Universidad Francisco de Vitoria, Ctra. M-515 Pozuelo-Maja-dahonda km. 1.800, 28223, Pozu...

work page 2018
[2]

This observation aligns with the bias-variance trade-off, as described by (Belkin et al., 2019)

MSE values for Autoencoders Dataset Training Validation Test Moving MNIST 2.78×10-5 3.43×10-5 4.20×10-5 ICPR’04 0.0010 0.0010 0.0015 UCF101 0.0017 0.0018 0.0018 11 All the results presented in Table 2 indicate that the models perform well, showing no signs of overfitting or underfitting. This observation aligns with the bias-variance trade-off, as describ...

work page 2019
[5]

2 Related works This paper aims to evaluate hybrid Deep Learning models for video frame predic-tion

Finally, Section 6 concludes the paper with a summary of key findings and directions for future research. 2 Related works This paper aims to evaluate hybrid Deep Learning models for video frame predic-tion. In particular, the proposed approaches benefit from other Deep Learning models and their capabilities to extract features from images and make an anal...

work page 2013
[6]

As can be seen in the previous Figure, the workflow implies different Deep Learning models that are defined as follows

Proposed workflow for video frame prediction based on feature maps. As can be seen in the previous Figure, the workflow implies different Deep Learning models that are defined as follows. Autoencoder. These are unsupervised learning models, training data is not labelled, originally introduced in (Ballard, 1987). A key characteristic of Autoencoders is tha...

work page 1987
[7]

excellent prediction

Mean SSIM values of the reconstructed feature maps test set images. Dataset Model Reconstructed Feature Maps Moving MNIST RNN 0.5804 LSTM 0.6289 GRU 0.6668 3D-CNN 0.7628 ConvLSTM 0.7518 RCNN 0.7079 ICPR’04 RNN 0.6133 LSTM 0.6419 GRU 0.6871 3D-CNN 0.8013 ConvLSTM 0.8081 RCNN 0.7132 14 UCF101 RNN 0.1618 LSTM 0.1931 GRU 0.1935 3D-CNN 0.5483 ConvLSTM 0.6981 R...

work page 1931
[9]

∑(𝑦#− ŷ#)%

𝑀𝑆𝐸= !"∑(𝑦#− ŷ#)%"#$! (2) For Equations 1 and 2 above, 𝑦# represents the value at a specific position in the ground truth image and ŷ# represents the corresponding value in the reconstructed image. The Structural Similarity Index Measure (SSIM), introduced by (Wang et al., 2004), assesses the perceptual similarity between images, regardless of their absol...

work page 2004
[10]

First, before training the models, the creation of the training, validation, and test subsets is needed

to obtain accurate models. First, before training the models, the creation of the training, validation, and test subsets is needed. In Deep Learning, it is common practice to split the data with an 80-20 ratio between training and test sets, followed by an 80-20 split of the training set to create the validation set. This approach ensures that each subset...

work page 2012
[11]

Inference analysis of the approaches. Dataset Approach Run time per iteration of pre-dictive models Total execution time of the approach Moving MNIST 3D-CNN 0.03 s 1,267 s + 1,079 s Baseline 0.21 s 6,990 s ICPR’04 ConvLSTM 0.06 s 706 s + 1,209 s Baseline 0.48 s 18,892 s UCF101 ConvLSTM 0.06 s 1,019 s + 1,038 s Baseline 0.32 s 14,455 s Analyzing Table 11, ...

work page 2025

Show all 16 references

[12]

This significant reduction trans-lates directly into a smaller carbon footprint and longer battery life when the models are deployed on embedded or edge devices

Even after accounting for the encoder and decoder phases, the feature-map strategy reduces the energy budget by 70 to 94%, an advantage which increases with input complexity: ICPR’04 and UCF101, which contain richer motion and higher spatial variability, benefit the most. This...

2024
[50]

https://doi.org/10.1007/s11554-025-01626-w Backus, M., Jiang, Y., & Murphy, D. (2022). Video Frame Prediction with Deep Learn-ing. Ballard, D. H. (1987). Modular learning in neural networks. AAAI, 647, 279–284. Belkin, M., Hsu, D., Ma, S., & Mandal, S. (2019). Reconciling mode...

2022
[846]

V., Al-Shehari, T., Alsadhan, N

https://doi.org/10.3390/s25030846 Aravinda, C. V., Al-Shehari, T., Alsadhan, N. A., Shetty, S., Padmajadevi, G., & Reddy, K. R. U. K. (2025). A novel hybrid architecture for video frame predic-tion: combining convolutional LSTM and 3D CNN. Journal of Real-Time Image Processing, 22(1),

2025 doi
[2014]

Regarding (Kemal Pola & Saban Öztürk, 2023), GRUs simplify the internal structure of LSTM cells by reducing the number of gates, thereby decreasing the model's time complexity

as a type of RNN designed to improve the computational efficiency of LSTM networks, especially when handling large-scale data. Regarding (Kemal Pola & Saban Öztürk, 2023), GRUs simplify the internal structure of LSTM cells by reducing the number of gates, thereby decreasing th...

2023
[2017]

∑∣𝑦#− ŷ#∣

is defined as a neural architecture that processes sequential input by feeding each frame of a sliding window into a recurrent layer. The model uses the outputs and hidden states of the recurrent units at each time step to extract temporal features across the sequence, enablin...

2019
[2021]

presents a novel next-frame video prediction method that combines MobileNetV2 for frame-level feature extraction with a Trans-former model for temporal sequence modelling of embeddings. Also (Mathai et al., 2024), which presents proposed HF2-VAD, a hybrid video anomaly detecti...

2024
[2022]

They find that the ConvLSTM model, which explicitly captures temporal and spatial patterns, outperforms GANs in predicting fu-ture frames

evaluate two Deep Learning architectures, Convolutional Long Short-Term Memory (ConvLSTM) and Generative Adversarial Networks (GANs), for human motion frame prediction, training them on the UCF101 dataset. They find that the ConvLSTM model, which explicitly captures temporal a...

2022
[2024]

as a discipline that seeks to equip computers with the ability to interpret, analyze, and understand visual information from their surround-ings, including images and videos. It is within computer vision that we find one of the great milestones of modern AI: the development of...

2012

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Sánchez Velázquez1, Mingbo Cai2, 3, Andrew Coney1, Álvaro J

Benefits of Feature Extraction and Temporal Sequence Analysis for Video Frame Prediction: An Evaluation of Hybrid Deep Learning Models Jose M. Sánchez Velázquez1, Mingbo Cai2, 3, Andrew Coney1, Álvaro J. García-Tejedor and Alberto Nogales1 1 CEIEC Research Institute, Universidad Francisco de Vitoria, Ctra. M-515 Pozuelo-Maja-dahonda km. 1.800, 28223, Pozu...

work page 2018

[2] [2]

This observation aligns with the bias-variance trade-off, as described by (Belkin et al., 2019)

MSE values for Autoencoders Dataset Training Validation Test Moving MNIST 2.78×10-5 3.43×10-5 4.20×10-5 ICPR’04 0.0010 0.0010 0.0015 UCF101 0.0017 0.0018 0.0018 11 All the results presented in Table 2 indicate that the models perform well, showing no signs of overfitting or underfitting. This observation aligns with the bias-variance trade-off, as describ...

work page 2019

[3] [5]

2 Related works This paper aims to evaluate hybrid Deep Learning models for video frame predic-tion

Finally, Section 6 concludes the paper with a summary of key findings and directions for future research. 2 Related works This paper aims to evaluate hybrid Deep Learning models for video frame predic-tion. In particular, the proposed approaches benefit from other Deep Learning models and their capabilities to extract features from images and make an anal...

work page 2013

[4] [6]

As can be seen in the previous Figure, the workflow implies different Deep Learning models that are defined as follows

Proposed workflow for video frame prediction based on feature maps. As can be seen in the previous Figure, the workflow implies different Deep Learning models that are defined as follows. Autoencoder. These are unsupervised learning models, training data is not labelled, originally introduced in (Ballard, 1987). A key characteristic of Autoencoders is tha...

work page 1987

[5] [7]

excellent prediction

Mean SSIM values of the reconstructed feature maps test set images. Dataset Model Reconstructed Feature Maps Moving MNIST RNN 0.5804 LSTM 0.6289 GRU 0.6668 3D-CNN 0.7628 ConvLSTM 0.7518 RCNN 0.7079 ICPR’04 RNN 0.6133 LSTM 0.6419 GRU 0.6871 3D-CNN 0.8013 ConvLSTM 0.8081 RCNN 0.7132 14 UCF101 RNN 0.1618 LSTM 0.1931 GRU 0.1935 3D-CNN 0.5483 ConvLSTM 0.6981 R...

work page 1931

[6] [9]

∑(𝑦#− ŷ#)%

𝑀𝑆𝐸= !"∑(𝑦#− ŷ#)%"#$! (2) For Equations 1 and 2 above, 𝑦# represents the value at a specific position in the ground truth image and ŷ# represents the corresponding value in the reconstructed image. The Structural Similarity Index Measure (SSIM), introduced by (Wang et al., 2004), assesses the perceptual similarity between images, regardless of their absol...

work page 2004

[7] [10]

First, before training the models, the creation of the training, validation, and test subsets is needed

to obtain accurate models. First, before training the models, the creation of the training, validation, and test subsets is needed. In Deep Learning, it is common practice to split the data with an 80-20 ratio between training and test sets, followed by an 80-20 split of the training set to create the validation set. This approach ensures that each subset...

work page 2012

[8] [11]

Inference analysis of the approaches. Dataset Approach Run time per iteration of pre-dictive models Total execution time of the approach Moving MNIST 3D-CNN 0.03 s 1,267 s + 1,079 s Baseline 0.21 s 6,990 s ICPR’04 ConvLSTM 0.06 s 706 s + 1,209 s Baseline 0.48 s 18,892 s UCF101 ConvLSTM 0.06 s 1,019 s + 1,038 s Baseline 0.32 s 14,455 s Analyzing Table 11, ...

work page 2025