This video presents a comprehensive comparison of nine popular 3D cameras, moving beyond technical specifications to highlight real-world image quality and practical issues. The cameras, including various Intel RealSense models, OAK-D Light and Pro, ZED Camera, Microsoft Azure Kinect, and the high-accuracy Mecademic 3D camera (used as a reference), are tested against a challenging setup featuring reflective metals, dark surfaces, repeating patterns, thin objects, and glass, under diverse lighting conditions. The analysis reveals specific strengths and weaknesses of each camera, such as stability issues, texture breakdown, handling of reflective or transparent surfaces, and performance under varying light. The video concludes by offering guidance on selecting the right 3D camera based on application-specific needs, distinguishing between general-purpose, industrial, prototyping, outdoor, and research-focused use cases.
Comparing 3D Cameras: A Deep Dive into Performance and Real-World Challenges
0:00 Introduction and Purpose: The video provides an epic comparison of nine popular 3D cameras, focusing on real-world image quality and practical problems rather than just technical specifications.
0:18 Cameras Included: The comparison features Intel RealSense (D455, D435, D450, D415), OAK-D Light, OAK-D Pro, ZED Camera, Microsoft Azure Kinect, and the Mecademic 3D camera (used as a high-accuracy reference).
0:11 Test Setup: Cameras are tested with challenging objects like shiny metal, dark items, repeating patterns, small/thin objects, glass, and a paper airplane to reveal common issues not found in benchmarks. A second scene covers close-range performance and challenging lighting (windows, occlusions).
0:03:17 Mecademic 3D Camera (Reference): Demonstrates the highest accuracy and minimal dispersion. It performs well on metal surfaces but struggles with black and dark, angled surfaces (e.g., the paper airplane's sides).
0:04:37 RealSense D455: Provides a nice image but with significantly lower accuracy and resolution than the Mecademic. It struggles with surfaces at small angles, black textures, and glass, though metal and paper generally work well. In the second scene, windows cause unpredictable artifacts and the image texture can be unstable.
0:07:07 RealSense D435: Similar to the D455, with the main difference being its narrower field of view, which significantly affects what's captured (e.g., the paper airplane is invisible in the first scene).
0:08:28 RealSense D450: Shows slightly better object quality but severely "destroys" the background plane, causing it to float with particles. This issue is unique among the RealSense cameras tested for this specific background.
0:09:46 RealSense D415: Optimized for close range and lacks active lighting. Its image quality is generally worse, with unstable and jittering 3D planes, leading to lower accuracy. Interestingly, it can sometimes see through windows.
0:11:07 OAK-D Light: Primarily focused on AI acceleration. Using a basic OpenCV 3D algorithm, it produces more artifacts than RealSense cameras but offers more stability than the D415. It handles metal better but struggles with white or structured objects like the background plane. It struggles with background/foreground due to different lighting conditions but surprisingly sees stably behind windows.
0:12:59 OAK-D Pro: Features active lighting and better resolution than the OAK-D Light. It exhibits a similar background plane instability problem to the RealSense D450. Active illumination helps it perform better than the OAK-D Light in dark conditions, though it has less impact on far objects.
0:14:11 Microsoft Azure Kinect: Equipped with larger lenses, it delivers better overall quality than RealSense or OAK-D cameras, showing fewer holes and more detail on objects like the paper airplane. However, its images can be unstable with flickering details, making its accuracy somewhat unreliable. It struggles with dark objects and glass. In the second scene, it shows good accuracy and stability, though with some noise behind windows.
0:16:38 ZED Camera: Popular and effective, it features a wider distance between lenses, which can affect close-range accuracy. It operates in two regimes: "Neural" (smoother image with noise/fluctuations in less defined areas) and "Classic" (more holes but lower noise/fluctuation), allowing users to choose based on their needs. It requires a more powerful host machine for its inference processes.
0:19:49 General Conclusions: All nine cameras are suitable for different applications. The choice depends on the underlying physics (stereo, active illumination), the host device, and specific camera characteristics like field of view, accuracy, frame rate, and size.