Hey! Let's break down these two core terms for you since you're diving into Structure-from-Motion Revisited—such a foundational paper for modern SFM pipelines. Here's what they mean in this context:

What is a "Track"? #

In SFM, a track refers to the linked sequence of matched feature points that all correspond to the same physical 3D point across multiple images. Think of it like this:

• Suppose you take 12 photos of a coffee mug from different angles.
• A tiny chip on the mug's rim is detected as a 2D feature point in 7 of those images, and the SFM algorithm successfully confirms these 7 points are all the same chip viewed from different perspectives.
• That connected set of 7 2D points (all mapping to the chip's 3D position) is a single track.

Tracks are the backbone of SFM: without establishing that multiple 2D points across images belong to the same 3D point, we can’t use triangulation to calculate the point’s 3D coordinates or estimate camera poses.

What is "Track Length"? #

Track length is straightforward—it’s the number of images (frames) that a given track appears in. Using the coffee mug example above, the chip’s track would have a length of 7.

This metric matters a lot for reconstruction quality:

• Longer tracks provide more geometric constraints from diverse camera viewpoints, leading to more accurate 3D positions and camera pose estimates.
• Short tracks (e.g., length 2) are often unreliable: they may come from weak feature matches or minimal perspective changes, which can introduce large errors. Many SFM pipelines filter out tracks below a minimum length (like 3 or 4) to boost overall result accuracy.

内容的提问来源于stack exchange，提问作者justPassBy