Athanasios Voulodimos, Nikolaos Doulamis, Anastasios Doulamis, Eftychios Protopapadakis
6 min
Abstract
Over the last years deep learning methods have been shown to outperform previous state-of-the-art machine learning techniques in several fields, with computer vision being one of the most prominent cases. This review paper provides a brief overview of some of the most significant deep learning schemes used in computer vision problems, that is, Convolutional Neural Networks, Deep Boltzmann Machines and Deep Belief Networks, and Stacked Denoising Autoencoders. A brief account of their history, structure, advantages, and limitations is given, followed by a description of their applications in various computer vision tasks, such as object detection, face recognition, action and activity recognition, and human pose estimation. Finally, a brief overview is given of future directions in designing deep learning schemes for computer vision problems and the challenges involved therein.
Alex: Greedy building avoids bad starting points without early error backprop. But for cluttered motion tracking, do they match CNNs?
Sam: They complement CNNs, often pretraining them to fight overfitting. Boltzmann models excel at generative tasks, reconstructing data for robust features, though training approximations make it tricky.
Alex: Less labeled data, but compute hurdles. Mixing them with CNNs seems smart for tough vision.
Sam: Spot on—unsupervised hierarchies from raw feeds set up better performance, outpacing handcrafted methods.
Alex: What makes greedy training in Deep Belief Networks great for initialization, and where do they falter on images?
Sam: It picks strong starting weights unsupervised, dodging poor optima. But high compute, unclear fine-tuning, and ignoring image layout hurt. Convolutional Deep Belief Networks fix that with convolutional RBMs, respecting pixel neighborhoods for translation-invariant handling of big images.
Alex: Spatial smarts like CNNs—perfect for visuals. How do Deep Boltzmann Machines differ?
Sam: All layers undirected, with alternating conditional ignores for efficiency. Inference is tough, so they approximate with stochastic methods after RBM pretraining, then joint fine-tune. Top-down feedback handles uncertainty well, like partial views in driving.
Alex: Useful for messy scenes, but why so compute-heavy?
Sam: Joint optimization on big data is brutal. Tweaks like separate initializers help, but scalability lags.
Alex: Now stacked autoencoders—how do they stack up as unsupervised blocks?
Sam: They compress input to a code, then reconstruct, learning nonlinear features like smart principal components. Denoising versions corrupt input—like scrambling pixels—but rebuild the clean one, capturing dependencies. Stack greedily, then fine-tune supervised.
Alex: Reconstruction-focused, not probabilistic sampling. Why do denoising ones flex better than plain Deep Belief Networks?
Sam: Smooth error functions allow flexible designs, unlike rigid RBM energies. But no true generation—no easy sampling to check distributions.
Alex: Head-to-head, CNNs top vision benchmarks, thanks to built-in invariance to shifts or scales—key for objects in any pose. DBNs skip labels but struggle with probabilities; autoencoders train fastest.
Sam: CNNs rule for that reason. Boltzmann models build label-free hierarchies; autoencoders enable real-time pretraining, especially denoising grabbing structure from noise.
Alex: Explains CNNs dominating pedestrian detection—identity holds despite pixel wiggles.
Sam: Object detection uses CNNs for proposals and classification into bounding boxes. Faces shifted to CNNs for invariance. Action recognition fuses them with motion models or LSTMs for video events. Pose estimation splits holistic—whole-image regression—or part-based, CNNs on patches linked for pose realism. Part-based often wins by nailing local details with heatmaps and consistency checks.
Alex: Deep models—from CNNs to probabilistics—swap rigid rules for adaptive learning, surging vision for robots and cars. CNNs lead with invariance but crave data; others add unsupervised flexibility, though compute and theory gaps persist. Thanks, Sam—this review nails why deep learning reshaped perception. Thanks for listening to ResearchPod.