Source-linked AI summary
Deep Learning vs. Traditional Computer Vision
Niall O' Mahony, Sean Campbell, Anderson Carvalho, Suman Harapanahalli, Gustavo Velasco-Hernandez, Lenka Krpalkova, Daniel Riordan, Joseph Walsh
TL;DR
The paper examines whether traditional computer vision remains necessary despite deep learning’s effectiveness. It compares both approaches and reviews hybrid methods, concluding that traditional techniques remain relevant for performance improvements and applications where deep learning is not yet well established.
Problem
The paper addresses whether deep learning has made traditional computer vision techniques obsolete and whether they should still be studied.
Method
The paper compares traditional computer vision and deep learning, then reviews hybrid approaches combining hand-crafted techniques with deep-learning models.
Results
Traditional computer vision remains relevant in hybrid methods and applications including SLAM, panoramic stitching, geometric deep learning, and 3D vision.
Takeaways & Limitations
Maintaining knowledge of traditional computer vision broadens the solutions available to computer-vision engineers, particularly for 3D problems and tasks where deep learning is overkill.
Takeaways & Limitations
Deep learning is not a panacea and has limited ability to learn visual relations such as whether multiple objects are the same or different.
Abstract
from arXiv · showhide
Deep Learning has pushed the limits of what was possible in the domain of Digital Image Processing. However, that is not to say that the traditional computer vision techniques which had been undergoing progressive development in years prior to the rise of DL have become obsolete. This paper will analyse the benefits and drawbacks of each approach. The aim of this paper is to promote a discussion on whether knowledge of classical computer vision techniques should be maintained. The paper will also explore how the two sides of computer vision can be combined. Several recent hybrid methodologies are reviewed which have demonstrated the ability to improve computer vision performance and to tackle problems not suited to Deep Learning. For example, combining traditional computer vision techniques with Deep Learning has been popular in emerging domains such as Panoramic Vision and 3D vision for which Deep Learning models have not yet been fully optimised
1 Introduction
The paper examines whether deep learning has made traditional computer vision obsolete and reviews how traditional methods can remain useful as alternatives, complements, and tools for problems deep learning cannot solve.
- 1 Introduction: Deep learning has substantially advanced digital image processing, but the paper questions whether its success makes traditional computer vision obsolete.The discussion is motivated by deep learning’s strong performance and dominance in computer vision since 2012.
- 1 Introduction: Traditional techniques may better address some computer-vision problems involving global features and challenges such as computing power, time, accuracy, and input constraints.The paper presents these techniques as possible ways to overcome challenges associated with deep learning.
- 1 Introduction: The paper compares deep learning with traditional hand-crafted feature approaches that dominated computer vision before deep learning.It reviews traditional algorithmic approaches and the applications where they can substitute for or complement deep learning.
- 1 Introduction: The paper reviews hybrid deep-learning and computer-vision methods, focusing on emerging 3D-perception tasks including object registration, detection, and point-cloud semantic segmentation.It also discusses directions for improving 3D deep learning and the broader impact of increased 3D use in computer vision.
2 A Comparison of Deep Learning and Traditional Computer Vision
Deep Learning improves accuracy, flexibility, and automation in computer vision, but traditional techniques remain useful where efficiency, transparency, limited resources, or non-class-specific processing matter. The paper compares both approaches and examines how they can be combined.
- Advantages of Deep Learning: Deep Learning improves accuracy in classification, semantic segmentation, object detection, and SLAM while requiring less manual feature engineering.Neural networks learn salient features from annotated data rather than relying on hand-crafted feature extraction.
- Traditional Computer Vision Workflow: Traditional object detection extracts descriptors such as SIFT, SURF, and BRIEF, then matches bag-of-words representations against new images.Feature extraction may include edge detection, corner detection, or threshold segmentation.
- Traditional Computer Vision Workflow: Traditional methods become cumbersome as classes increase because engineers must select informative features and fine-tune many parameters.This workflow depends heavily on expert judgment and trial and error.
- Advantages of Deep Learning: Deep Learning automates feature discovery through end-to-end training, but trades reduced manual engineering for substantial computing requirements and training time.The workflow shifts expertise toward iterating through deep learning architectures.
- CNN Workflow: CNNs use trained kernels to detect image features, while pooling reduces feature redundancy before later layers produce representations and output scores.The paper presents convolutional and pooling layers as core CNN building blocks.
- Advantages of Traditional Computer Vision Techniques: Traditional techniques can be more efficient, transparent, and general for closed applications, image stitching, and 3D mesh reconstruction that do not require class-specific knowledge.They can also support low-cost deployment, data augmentation, dataset annotation, and difficult vision problems such as robotics and 3D modelling.
3 Challenges for Traditional Computer Vision
Traditional computer vision and Deep Learning involve different trade-offs in transparency, efficiency, accuracy, versatility, and resource demands. Hybrid systems combine them to improve performance and reduce computational requirements.
- Trade-offs: Classic computer vision is transparent, well-established, and optimized for performance and power efficiency.These properties distinguish it from approaches that depend on larger computational resources.
- Trade-offs: Deep Learning offers greater accuracy and versatility but requires large amounts of computing resources.The paper frames this as a direct trade-off with the efficiency of classic algorithms.
- Hybrid Approaches: Hybrid systems can detect faces, features, or moving objects with computer vision before applying a DNN to a smaller image patch for verification or classification.This divides processing between efficient traditional detection and learned recognition.
- Hybrid Approaches: 130X-1,000X reduction in multiply-accumulate operations and about 10X improvement in frame rates are reported for hybrid implementations versus pure DL solutions.The hybrid implementation also uses about half the memory bandwidth and significantly lower CPU resources.
3.2 Overcoming the Challenges of Deep Learning
Deep learning faces substantial computing, data, resolution, and training-time demands. Traditional computer vision and preprocessing can reduce these demands or support performance when data and computing resources are limited.
- DL requires substantial processing power, including dedicated hardware, because higher accuracy can require billions of additional math operations.
- High-resolution inputs improve some DL vision tasks but increase the data, storage, transfer, and processing requirements.Traditional features or optical flow can first identify regions of interest, reducing training time and data requirements.
- Large datasets are often needed for DL, while traditional methods remain useful when big datasets or high-computing facilities are unavailable.Examples include ImageNet with 1.5 million images and COCO with 2.5 million images.
- DNN training can take hours or days and often requires many iterations with different training parameters.Transfer learning is identified as a common way to reduce training time.
- Domain-specific tasks may require hundreds or thousands rather than millions of examples, and traditional CV can augment or preprocess limited data.Preprocessing can make relationships easier for the model to interpret, while augmentation uses transformations such as rotations and shifts.
3.3 Making Best Use of Edge Computing
Edge computing benefits from combining conventional computer vision with deep learning across heterogeneous hardware. Such hybrids can reduce resource demands while preserving learned-model capabilities.
- Edge execution can reduce latency, cloud storage and processing, and bandwidth requirements compared with cloud-based implementations.Avoiding network transmission can also support privacy and security requirements for sensitive data.
- Hybrid CV-DL systems assign different workloads across CPUs, DSPs, FPGAs, coprocessors, and AI accelerators to improve power efficiency.The approach exploits heterogeneous computing capabilities available at the edge.
- Hybrid methods combine hand-crafted features with learned features or combine linear subspace methods with deep convolutional prediction.These approaches are reported for facial-expression recognition and 3D robot vision.
- Combining linear subspace methods with deep convolutional prediction achieved improved performance and several orders of magnitude faster runtime than the state of the art.
3.4 Problems Not Suited to Deep Learning
Several computer-vision problems remain poorly suited to deep learning because they are difficult to implement differentiably, require strong priors, or depart from DL’s usual data assumptions. Traditional techniques therefore retain a role in these settings.
- Robotics, panorama stitching, augmented and virtual reality, 3D modelling, motion capture, video processing, and scene understanding may require traditional techniques.These problems cannot simply be implemented in a differentiable manner with deep learning.
- DL excels at closed-end classification when sufficient data are available and test data resemble training data, but deviations from these assumptions can cause problems.
- DL has limited ability to learn visual relations, such as whether multiple objects in an image are the same or different.Feedback mechanisms including attention and perceptual grouping are suggested as possible computational components for abstract visual reasoning.
- Machine-learning models may require injected priors because not everything can be learned from data.3D computer vision can require priors such as smoothness, silhouette, and illumination information.
- Emerging computer-vision fields present new challenges for DL and may give classic computer vision a more prominent role.
3.5 3D Vision
3D vision extends deep learning into volumetric and non-Euclidean data, but larger inputs, cubic computational growth, occlusions, and noisy scans create major challenges. Hybrid architectures, preprocessing, and hand-crafted geometric features are used to address them.
- 3D CNNs process 3D vision data, preserving spatial information through kernels whose depth is smaller than the input volume.The paper situates these models within geometric deep learning and applications including video, graphics, vision, and robotics.
- 3D CNN computational complexity grows cubically with resolution because inputs require more memory and convolution across three dimensions.Occlusions and different camera angles add uncertainty compared with 2D image processing.
- 3D object detection in point clouds is challenging because discrete sampling, noisy scans, occlusions, and cluttered scenes complicate perception.
- FFT-based methods can reduce 3D CNN computation but increase memory requirements, while WMFA achieved a twofold speedup over cuDNN without increasing required memory.
- Euclidean 3D data can use existing 2D CNN paradigms, whereas non-Euclidean data lack a global grid and are harder to process with classical DL extensions.Voxel approaches are described as suitable for simple rigid objects, but voxelization can lose continuous shape information.
- A rotation-invariant mean-curvature feature increased Octnet accuracy by 1% overall on ModelNet10.
3.6 SLAM
Visual SLAM uses vision instead of LiDAR to register scene landmarks, combining sensing, matching, motion estimation, and landmark updates. In image-based localization, global descriptors provide coarse localization before geometric pose refinement, while 3D feature extraction remains challenging.
- Visual SLAM replaces LiDAR with vision for landmark registration while offering rich data, low cost, lightweight hardware, and lower post-processing workload.Its workflow includes environment sensing, data matching, motion estimation, location updates, and registration of new landmarks.
- Zero/one-shot learning under changes in rotation, scale, and lighting remains challenging, making feature extraction and representation useful for reducing training examples.
- Image-based localization commonly performs place recognition first and pose estimation second.Place recognition retrieves a database image with a similar global descriptor for coarse localization; pose estimation then refines the query pose using PnP and geometric verification.
- LiDAR scans lack a local-feature extraction algorithm comparable to SIFT, motivating approaches that transfer or construct descriptors from associated image and depth information.One approach associates 3D points with SIFT descriptors from triangulated images, while alternatives use RGB-D features and traditional descriptors.
3.7 360 cameras
360 cameras provide near-spherical visual coverage that can replace multiple monocular cameras and reduce blind spots, but fisheye distortion complicates conventional vision methods. Hybrid processing combines geometric reasoning with DL-based abstraction for indoor-scene layout hypotheses.
- 360 cameras capture approximately the entire sphere, enabling broad visual coverage and eliminating blind spots in robotic vehicles.Their 360-degree field of view can replace multiple monocular cameras and removes limitations on steering choices.
- Heavy barrel distortion from ultra-wide-angle fisheye lenses complicates conventional lane-detection and trajectory-tracking methods.Calibration and dewarping are often required as additional preprocessing steps.
- Hybrid scene understanding combines geometric reasoning from lines and vanishing points with DL-based abstraction from edge and normal maps.The combination is used to extract structure and generate layout hypotheses for indoor scenes.
- In sparsely structured scenes, direct image alignment can replace feature-based alignment when distinct image features are scarce.Phase-correlation methods are one option, while DCF-based correlation methods are reported to outperform phase-correlation approaches on the cited datasets.
3.8 Dataset Annotation and Augmentation
The paper identifies data scarcity and annotation effort as major constraints for specialized DL applications, especially pixel-level semantic segmentation. Traditional CV supports dataset creation and augmentation to reduce these burdens.
- Specialized DL applications face insufficient data, while semantic segmentation makes annotation especially burdensome because every pixel requires accurate labeling.The paper identifies dataset annotation as a major DL bottleneck requiring many hours of manual work.
- Traditional algorithmic tools can semi-automate annotation using ORB features, polygon morphing, and semi-automatic Area of Interest fitting.
- Dataset augmentation enlarges limited image-classification datasets through label-preserving transformations such as cropping, scaling, and rotation.The process is intended to reduce overfitting while remaining computationally light and compatible with DL training pipelines.
- Traditional augmentation methods include PCA, noise addition, feature-space interpolation or extrapolation, and visual-context modeling from segmentation annotations.
Conclusion
The conclusion argues that traditional CV remains useful alongside DL, particularly as an alternative for tasks where DL is excessive and as a component of hybrid methods. It highlights applications where traditional techniques can reduce requirements or support emerging 3D and panoramic systems.
- Traditional CV remains useful even in the DL era and can provide additional intuitions and tools, especially for 3D computer-vision problems.The paper warns that relying only on DL limits the range of solutions available to CV engineers.
- Traditional CV can serve as an alternative to DL when DL is overkill for a specific task.
- Hybrid approaches use traditional CV to improve performance and reduce training time, processing, and data requirements.
- Traditional techniques remain relevant in SLAM, panoramic stitching, geometric deep learning, and 3D vision, where DL is not yet well established.
- The paper reviews cases showing that pre-DL CV techniques can still provide value after the field’s shift toward data-driven intelligence.