litert-community/Ultra-Fast-Lane-Detection-LiteRT

Ultra-Fast-Lane-Detection (ResNet18, CULane) — LiteRT GPU

On-device lane detection running fully on the LiteRT CompiledModel GPU delegate (no CPU fallback). Ultra-Fast-Lane-Detection (ECCV 2020) reformulates lane detection as fast row-wise classification — the network runs on the GPU, and a tiny host-side arg/expectation decode turns the grid into lane points. ~20 ms/frame on a Pixel 8a.

• Architecture: ResNet18 backbone + row-anchor classification head — pure CNN.
• Weights: cfzd/Ultra-Fast-Lane-Detection (CULane, ResNet18) · MIT.
• Size: 178 MB.

Detected ego-lane on a dashcam highway frame. Source: Wikimedia Commons (Public Domain).

I/O

• Input: [1, 3, 288, 800] NCHW, RGB, x/255 then ImageNet-normalized (mean [0.485,0.456,0.406], std [0.229,0.224,0.225]).
• Output: [1, 201, 18, 4] = (griding+1, row_anchors, lanes) — per-lane, per-row classification logits over 200 horizontal grid cells (+1 "no lane").

Host-side decode

For each of the 4 lanes and 18 row anchors: softmax over the 200 grid cells, take the expectation → column; if the argmax over all 201 is the last index (200 = "no lane"), drop it. Map the column to an x-pixel via linspace(0, 799, 200) (scaled to the image width) and the row anchor to a y-pixel (CULane row anchors, scaled from 288).

GPU conversion

UFLD is a pure CNN. It converts fully GPU-compatible (41/41 nodes on the delegate, 1 partition; device corr 0.999982, ~20 ms) with one patch: the ResNet18 stem MaxPool2d(padding=1) lowers to a -inf PADV2 (rejected by Mali), replaced by a 0-pad

• unpadded maxpool (exact post-ReLU). CPU-exact vs PyTorch (corr 0.9999999999996).

Minimal usage

Kotlin (Android, LiteRT CompiledModel GPU)

val options = CompiledModel.Options(Accelerator.GPU)
val model = CompiledModel.create(context.assets, "ufld.tflite", options, null)
val inBufs = model.createInputBuffers()
val outBufs = model.createOutputBuffers()

inBufs[0].writeFloat(inputNCHW)          // [1,3,288,800] RGB, x/255 then ImageNet-norm
model.run(inBufs, outBufs)
val out = outBufs[0].readFloat()         // [201*18*4], layout (griding+1, rows, lanes)
// decode: per (lane,row) softmax over the first 200 cells, take the expectation -> column;
// skip if argmax == 200 (no lane). See LaneDetector.kt for the full decode.

Python (LiteRT / ai-edge-litert)

import numpy as np
from ai_edge_litert.interpreter import Interpreter

it = Interpreter(model_path="ufld.tflite"); it.allocate_tensors()
inp, out = it.get_input_details(), it.get_output_details()
it.set_tensor(inp[0]["index"], x)        # [1,3,288,800] float32, RGB /255, ImageNet-norm
it.invoke()
o = it.get_tensor(out[0]["index"])[0]    # [201,18,4]
o = o[:, ::-1, :]
prob = np.exp(o[:-1]) / np.exp(o[:-1]).sum(0, keepdims=True)
loc = (prob * (np.arange(200) + 1).reshape(-1, 1, 1)).sum(0)   # [18,4] columns
loc[np.argmax(o, 0) == 200] = 0                                # 0 = no lane

Conversion

Converted with litert-torch (build_ufld.py): loads the ResNet18 CULane weights and exports the row-classification graph.

Performance

Measured on a Pixel 8a (Tensor G3, Android 16) with the standard TFLite benchmark_model tool — 10 warm-up runs then 50 timed runs, reported as the tool's mean.

Runtime	Backend	Graph on GPU	Latency
LiteRT CompiledModel (LITERT_CL)	GPU	41 / 41	~20 ms
TFLite benchmark_model (TfLiteGpuDelegateV2)	GPU (OpenCL)	41 / 41	25.3 ms
TFLite benchmark_model	CPU (XNNPACK, 4 threads)	—	252.3 ms

The two GPU rows are different runtimes, not a contradiction. The LITERT_CL figure is the one recorded when this model shipped, taken through LiteRT's own CompiledModel accelerator — the path the Kotlin sample app and the LiteRT API use. The TfLiteGpuDelegateV2 figure is the classic TFLite OpenCL delegate, measured with a tool anyone can download and re-run. They agree on how much of the graph the GPU takes; they disagree on speed, and the classic delegate is the slower of the two here. Read the TfLiteGpuDelegateV2 row as a reproducible floor, not as this model's speed on LiteRT.

Snapdragon NPU (Hexagon)

The NPU is 2.20x faster than the GPU (2.96 ms against 6.51 ms) and loads 7.44x faster (113 ms against 840 ms).

backend	inference (median / min)	load
NPU (Hexagon v81)	2.96 ms / 2.92 ms	113 ms
GPU (Adreno)	6.51 ms / 6.22 ms	840 ms

Measured on a Samsung Galaxy S26 (Snapdragon 8 Elite Gen 5 / SM8850, Hexagon v81, Android 16), LiteRT CompiledModel 2.2.0, one accelerator per process, 5 warm-up runs then N=50 timed runs, median reported. Every run held thermal status NONE throughout. Headroom 0.66-0.67, where 1.0 is the throttling threshold.

The NPU rows here ran artifacts compiled ahead of time for SM8850 with QAIRT 2.47.0; the GPU rows ran the published files as they are. LiteRT can also compile for the NPU on the device at first load, which is what lets you ship the published file unchanged — that path and the ten runtime libraries it needs are in the NPU recipe, and we did not measure it here. GPU wiring is in the GPU recipe.

License

MIT (Ultra-Fast-Lane-Detection / cfzd). Trained on CULane.