| """Phase 2: distance to an object from one RGB image, by geometry. |
| |
| No depth sensor and no learned depth model. Everything here works from focal |
| length, principal point, camera height and the pixel coordinates of a box. The |
| LiDAR-derived `gt_distance_m` in the manifest is used to score the result and |
| never appears in the estimation path -- delete `evaluate.py` and the rest still |
| runs. |
| |
| Transcription order |
| ------------------- |
| Each file only depends on the ones above it, so read and write them in this |
| order and you can run what you have at every step. |
| |
| 1. `ground_plane.py` The idea. A pixel is a ray; a ray that points below the |
| horizon hits the road; that intersection is a distance. |
| `ground_depth` is the whole thing in a dozen lines and |
| everything else is scaffolding around it. Start here. |
| |
| 2. `known_size.py` The other idea, and the shorter one. If you know how big |
| something really is, how big it looks tells you how far |
| away it is. Three functions, all one-liners. |
| |
| 3. `estimate.py` Where the two meet a real detection box. Which estimator |
| a class gets, how to sample a box rather than trusting |
| one pixel, and what to return when neither works. |
| |
| 4. `evaluate.py` Scoring against LiDAR truth. Nothing conceptual, but the |
| slicing in `report` is what made the per-scene error |
| visible, and that finding mattered more than any of the |
| estimator code. |
| |
| Four things worth having in mind before you start |
| ------------------------------------------------- |
| * **Distance means forward distance.** The Z component in the camera frame, |
| not Euclidean range. It has to match `manifest.gt_distance_m` or the scores |
| are meaningless. An object 30 m ahead and 10 m left is at 30, not 31.6. |
| |
| * **Camera height means height above the road.** nuScenes puts its ego origin |
| on the ground so its translation's z component is the height; Argoverse 2 |
| puts it at the rear axle, ~0.26 m up, and using the translation there makes |
| every distance read ~18% short. `Calibration.cam_height_m` is corrected. |
| |
| * **A pixel is not a pinhole pixel until you undistort it.** Argoverse 2 |
| cameras carry k1 near -0.24 on unrectified images. nuScenes ships rectified |
| frames and passes zeros, so the two paths only agree once distortion is in |
| the model. |
| |
| * **The bottom edge of the box is the whole estimate.** Everything above it is |
| in the air, and a ray through a point in the air lands on the road somewhere |
| well beyond the object. The most common way to get a plausible-looking wrong |
| answer is to sample the middle of a box. |
| |
| What the numbers ended up saying |
| -------------------------------- |
| Sub-metre under 10 m, and error growing as Z^2 after that -- which is inherent, |
| since sensitivity is `Z^2 / (fy * h)` and no estimator recovers information the |
| pixel does not contain. But the *dominant* error term turned out not to be range |
| at all: it is a per-scene offset, running -35% to +35% between scenes and close |
| to constant within one, which a single constant row shift per scene halves. That |
| is the flat-road assumption meeting a gradient. Per-frame ground-plane estimation |
| is worth more than any refinement of the maths in these four files. |
| """ |
|
|
