Source-linked AI summary
An Ensemble-Based Self-Taught Learning Approach for Parking Space Classification Under Limited Data
Lucas de Oliveira Cunha, Joelton Deonei Gotz, Paulo Lisboa de Almeida, Andre Gustavo Hochuli
TL;DR
Parking classification systems often need extensive target-domain labels and may generalize poorly across heterogeneous environments. This paper learns representations from unlabeled source data with heterogeneous autoencoders, then trains fixed-encoder target classifiers with limited labels and fuses their predictions. Under cross-dataset evaluation, it reports competitive accuracy across target scenarios with substantially reduced annotation requirements.
Problem
Parking classifiers often rely on large annotated datasets and have limited robustness across heterogeneous environments, while manual target-lot labeling is costly.
Source: §I · p. 1Method
An ensemble of heterogeneous convolutional autoencoders learns from unlabeled source data; frozen encoders support limited-label target classifiers whose posterior predictions are fused.
Source: §III · p. 2Results
With 64 annotated samples, average accuracy is about 93%, and with 1,024 samples it is approximately 96%, within two percentage points of Santos et al.’s GAN-based method.
Source: C. State-of-Art Comparison · p. 5Takeaways & Limitations
The framework achieves competitive performance across twelve target scenarios with limited supervision and generalizes without additional weight optimization.
Source: §V · p. 6
Abstract
from arXiv · showhide
Parking spot classification is a fundamental task in intelligent transportation systems, yet most deep learning approaches rely on large amounts of annotated data and exhibit limited generalization across heterogeneous environments. To address these limitations, we investigate a self-taught learning framework based on unsupervised representation learning with convolutional autoencoders. The proposed approach learns transferable visual representations from unlabeled data and reuses the learned encoders as fixed feature extractors for supervised classification with limited annotated samples in the target domain. To further enhance robustness and mitigate architectural bias, an ensemble of heterogeneous autoencoders is employed, with independent classifier heads and prediction fusion at inference time. Experiments conducted on the PKLot and CNRPark benchmarks under cross-dataset evaluation protocols show that the proposed ensemble-based strategy substantially reduces annotation requirements while improving robustness under significant domain shifts, achieving accuracies between 93\% and 96\% in data-constrained scenarios.
I. INTRODUCTION
Parking classification supports transportation tasks, but annotated-data costs and weak cross-environment robustness limit many deep-learning approaches. This work applies self-taught learning with heterogeneous autoencoders and evaluates it under cross-dataset shifts and limited target labels.
I. INTRODUCTION
- Most existing approaches depend on large scenario-specific labeled datasets and show limited robustness across heterogeneous parking environments.Performance can degrade under unseen conditions.Source: §I · p. 1
- Manual target-lot annotation is costly and time-consuming, often requiring thousands of occupancy-labeled images and motivating methods that need fewer labels.Source: §I · p. 1
- Self-Taught Learning transfers representations learned from unlabeled data to target tasks with limited labeled samples, and this work applies it using heterogeneous convolutional autoencoders.The ensemble is proposed to learn complementary representations for cross-dataset robustness and generalization.Source: §I · p. 1
- The evaluation on PKLot and CNRPark varies source domains and target-label counts to assess annotation reduction, domain-shift robustness, and comparison with fully supervised methods.Source: §I · p. 1
II. STATE OF THE ART
Prior parking-classification work includes large labeled datasets, cross-view and cross-domain models, and annotation-saving strategies. The review identifies limited evaluation under domain shift with scarce target labels and positions self-taught learning as unexplored for parking classification.
II. STATE OF THE ART
- PKLot contains about 700,000 manually annotated images; its original method exceeded 99% accuracy with extensive target-lot labels but fell to around 90% on unseen lots without domain adaptation.Source: §II · pp. 1–2
- Cross-view results include 88.5% accuracy for Amato et al. and 96.4% for Nurullayev and Lee, while a global PKLot/CNRPark-EXT model reached 92.8% on unseen lots.Source: §II · p. 2
- Bounding-box annotations can reduce effort while reaching up to 97% accuracy with 1,000 labels, and the review limits comparison to methods evaluated on public datasets.Source: §II · pp. 1–2
- Annotation-saving alternatives include teacher–student distillation at 96.6% and GAN-based synthesis at about 97% with 256 target labels, though the latter uses a complex, costly training pipeline.Source: §II · p. 2
- Self-taught learning reuses unsupervised representations as fixed feature extractors, but its effectiveness for parking classification had remained unexplored.Most existing methods assume closed-domain evaluation, with few addressing adaptation under scarce labeled target data.Source: §II · pp. 1–2
III. PROPOSED METHOD
The proposed method combines heterogeneous convolutional autoencoders to learn transferable visual representations without labels, then adapts them to target parking scenarios using limited labeled data. It also examines source-domain choice, label reduction, and generalization.
III. PROPOSED METHOD
- Heterogeneous convolutional autoencoders learn complementary representations from unlabeled data, which are reused to train lightweight target classifiers with limited labels.Architectural diversity is intended to mitigate bias and improve robustness to domain shifts.Source: §III · p. 2
- The framework trains autoencoders independently on unlabeled source data, freezes their encoders as feature extractors, trains lightweight classifiers on target labels, and fuses posterior probabilities at inference.Source: §III · p. 2
- The evaluation varies source-domain similarity and target-label availability, and assesses the approach’s generalization capability.Source: §III · p. 2
A. Representation Learning
Representation learning uses independently trained convolutional autoencoders to capture diverse visual features from unlabeled data. A stochastic generator varies their architecture and latent dimensions, while each decoder mirrors its encoder for reconstruction.
A. Representation Learning
- Each convolutional autoencoder uses an encoder with progressive spatial downsampling and a symmetric decoder for reconstruction.Source: A. Representation Learning · p. 2
- A stochastic generator creates heterogeneous encoder architectures with varied depth, downsampling strategies, and latent dimensionality.The architecture guidelines follow Delazeri et al..Source: A. Representation Learning · p. 3
- The generator samples two to five convolutional blocks, filter counts, pooling operations, and a latent dimension from 128 to 2,048.Pooling is limited to four operations for 64 × 64 inputs.Source: A. Representation Learning · p. 3
- The decoder expands the latent vector and uses transposed convolutions to reconstruct the input, with stride choices corresponding to encoder pooling stages.Its final layer uses a 3 × 3 convolution with sigmoid activation for normalized pixel intensities.Source: A. Representation Learning · p. 3
B. Task Learning
Task learning reuses each learned encoder as a frozen feature extractor and trains a lightweight classifier on its latent features.
B. Task Learning
- The learned encoder is reused as a fixed convolutional feature extractor, with all parameters frozen.Freezing preserves the learned representations.Source: B. Task Learning · p. 3
- A lightweight classifier head applies dropout at rate 0.2, then two ReLU fully connected layers with 256 and 128 neurons.Source: B. Task Learning · p. 3
- The classifier ends with two Softmax output units that produce posterior class probabilities.Source: B. Task Learning · p. 3
C. Training and Inference
Each autoencoder is independently trained on unlabeled source data, while lightweight classifier heads learn from extracted target-domain features and their predictions are combined at inference.
C. Training and Inference
- Each CAE is independently optimized on unlabeled source data by minimizing mean squared reconstruction error.Source: C. Training and Inference · p. 3
- The autoencoders are not parameter-shared or jointly optimized, promoting different latent representations across the ensemble.Source: C. Training and Inference · pp. 3–4
- A lightweight classifier is trained on extracted latent features, and CAE–classifier predictions are combined by majority voting for the final class decision.Source: C. Training and Inference · p. 4
D. Datasets
The study uses PKLot and CNRPark-EXT, two benchmarks with real-world environmental variability and distinct scenario or camera-view coverage.
D. Datasets
- The paper uses PKLot and CNRPark-EXT as vision-based parking-occupancy benchmarks, and Table I summarizes their properties.Source: D. Datasets · p. 4, Table I · p. 4
- PKLot contains 12,417 images and approximately 700,000 manually labeled parking spaces across UFPR04, UFPR05, and PUCPR.Source: D. Datasets · p. 4
- CNRPark-EXT includes approximately 160,000 labeled parking spaces from nine static viewpoints at one lot, with illumination artifacts, raindrops, and partial occlusions.Source: D. Datasets · p. 4
IV. EXPERIMENTS
The experiments use a cross-dataset protocol with strict separation between representation-learning and target-adaptation data. Unsupervised representations are trained from sampled images, and reconstructions are assessed qualitatively.
IV. EXPERIMENTS
- The cross-dataset protocol separates representation-learning datasets from target-domain adaptation data and treats each target camera as an independent scenario, yielding 12 scenarios.Source: §IV · p. 4
A. Unsupervised Representations Learning (STEP 1)
- Ten autoencoders were trained independently for 50 epochs, with each selected by its lowest validation MSE after training on 1,000 sampled images and 64 validation images.Source: A. Unsupervised Representations… · p. 4
- Reconstruction examples compare real samples with reconstructed ones; despite some geometric distortions, the stated goal is exposure to diverse shapes and textures relevant to occupancy cues rather than vehicle identity.Source: A. Unsupervised Representations… · p. 4, Figure 5 · p. 4
B. Supervised Task Learning (STEP 2)
The fixed encoders support supervised target-domain classification with limited labeled samples, and evaluation reports transfer across camera scenarios with diminishing gains from larger ensembles.
B. Supervised Task Learning (STEP 2)
- Each selected encoder remains frozen as a feature extractor, paired with an independently trained classifier head.Each classifier is trained for 10 epochs, and the model with the lowest validation loss is selected.Source: B. Supervised Task Learning (STEP 2) · pp. 4–5
- Representative parking spots across PKLot and CNRPark-EXT camera views illustrate the domain shifts considered in evaluation.The passage identifies the examples but does not quantify their differences.Source: B. Supervised Task Learning (STEP 2) · p. 5, Figure 6 · p. 5
- The target-domain training sets contain 64, 128, 256, 512, or 1,024 chronologically collected labeled samples, excluded from testing.The sample counts are used only in supervised task learning.Source: B. Supervised Task Learning (STEP 2) · p. 5
- Accuracy improves with more annotated target samples, with several scenarios performing strongly at 64–128 samples and gains usually saturating beyond 256.More challenging camera views are more sensitive to annotation size.Source: Table II · p. 5, B. Supervised Task Learning (STEP 2) · p. 5
- Competitive accuracy is maintained in most cross-domain settings, although transfers between datasets with markedly different acquisition conditions show more degradation.The framework generalizes without additional weight optimization.Source: B. Supervised Task Learning (STEP 2) · p. 5
- Cross-domain generalization averaged across scenarios increases with ensemble size but plateaus around 10 autoencoders.Improvements beyond 10 models are negligible, which the authors describe as a favorable cost-benefit trade-off.Source: Figure 7 · p. 5, B. Supervised Task Learning (STEP 2) · p. 5
C. State-of-Art Comparison
The comparison reports competitive cross-dataset accuracy with few labeled target samples, while contrasting annotation needs and deployment training with prior methods.
C. State-of-Art Comparison
- At 1,024 annotated target samples, average accuracy is approximately 96%, within two percentage points of Santos et al..Deployment requires training only the classifier head.Source: C. State-of-Art Comparison · p. 5
- Compared with Alves et al., the method avoids several days of target-data collection and full classifier fine-tuning, but still requires manually labeled samples.Alves et al. rely on pseudo-labels from the target parking lot.Source: C. State-of-Art Comparison · pp. 5–6
- About 93% average accuracy with 64 annotated samples outperforms several approaches in the state-of-the-art comparison.The comparison focuses on cross-dataset evaluation, target-domain supervision, and annotation requirements.Source: C. State-of-Art Comparison · p. 5, Table IV · p. 6
V. CONCLUSION
The conclusion presents heterogeneous convolutional autoencoders as a way to learn representations from unlabeled data and adapt them with limited supervision. Cross-dataset experiments report competitive performance across target scenarios.
V. CONCLUSION
- The framework combines unsupervised representation learning with efficient task adaptation for parking spot occupancy classification.It trains heterogeneous convolutional autoencoders on unlabeled data and reuses their encoders as fixed feature extractors.Source: §V · p. 6
- Cross-dataset experiments report competitive performance across 12 target scenarios, including with as few as 64 annotated samples.Performance saturates with modest annotation effort, and cross-scenario evaluations show generalization without additional weight optimization.Source: §V · p. 6
- The authors characterize the framework as an effective and scalable alternative to fully supervised methods for real-world parking monitoring.This conclusion is based on the reported benchmark and cross-dataset evaluations.Source: §V · p. 6