litert-community/YOLACT-ResNet50-LiteRT

Measured on device (edge-compat): Galaxy S26 · LiteRT 2.2.0 · GPU (ML Drift) · 38.4 ms p50 (2026-08-26); Galaxy S26 · LiteRT 2.2.0 · NPU (QNN/HTP) · 14.3 ms p50 (2026-08-26); Raspberry Pi 5 · LiteRT 2.2.0.dev20260804 · CPU/XNNPACK, 4 threads · 1096 ms p50 (2026-08-31); browser · Chromium 151 on M4 Max · LiteRT.js 2.5.3 · WebGPU · 63.1 ms p50 · output matches CPU (2026-08-11). Record: https://github.com/john-rocky/edge-compat/blob/main/cards/yolact-resnet50/CARD.md

YOLACT-ResNet50 — LiteRT (real-time instance segmentation, GPU)

On-device real-time instance segmentation running fully on the LiteRT CompiledModel GPU delegate (no CPU fallback). YOLACT (ICCV 2019) predicts per-instance COCO masks. The network (ResNet50 + FPN + protonet + heads) runs on the GPU; the lightweight decode (NMS + linear-combination masks) runs host-side. ~41 ms/graph on a Pixel 8a.

• Architecture: YOLACT-ResNet50 (base, no deformable conv) — pure CNN.
• Weights: dbolya/yolact (yolact_resnet50_54_800000) · MIT.
• Size: 125 MB.

Files

• yolact.tflite — the GPU graph (input [1,3,550,550] NCHW).
• priors.bin — 19248 SSD priors [cx,cy,w,h] (float32) used by the host-side box decode.

I/O

• Input: [1, 3, 550, 550] NCHW, BGR, normalized (x - [103.94,116.78,123.68]) / [57.38,57.12,58.40] (no /255).
• Raw outputs: loc [1,19248,4], conf [1,19248,81] (softmax, incl. background), mask [1,19248,32] (coefficients), proto [1,138,138,32] (prototype masks).

Host-side decode

1. Boxes: SSD decode(loc, priors, variances=[0.1,0.2]).
2. NMS: per-class, score-threshold ~0.3, IoU 0.5, top-k.
3. Masks (lincomb): for each kept detection, mask = sigmoid(proto @ coeff) → crop to the box → threshold 0.5 → upscale.

GPU conversion

Base YOLACT is a pure CNN, so the graph converts fully GPU-compatible (138/138 nodes on the delegate, 1 partition; device corr 0.99999–1.0 vs PyTorch on all four raw outputs) with one patch: the ResNet50 stem MaxPool2d(padding=1) lowers to a -inf PADV2 (rejected by Mali), replaced by a 0-pad + unpadded maxpool (exact post-ReLU). The scripted FPN is made traceable by disabling YOLACT's JIT (use_jit=False). CPU-exact vs PyTorch (corr 1.0).

Minimal usage

Kotlin (Android, LiteRT CompiledModel GPU)

val options = CompiledModel.Options(Accelerator.GPU)
val model = CompiledModel.create(context.assets, "yolact.tflite", options, null)
val inBufs = model.createInputBuffers()
val outBufs = model.createOutputBuffers()   // map by size: loc=N*4, conf=N*81, mask=N*32, proto=138*138*32

inBufs[0].writeFloat(inputNCHW)              // [1,3,550,550] BGR, (x-[103.94,116.78,123.68])/[57.38,57.12,58.40]
model.run(inBufs, outBufs)
val loc = outBufs[iLoc].readFloat()          // [19248*4]
val conf = outBufs[iConf].readFloat()        // [19248*81] (softmax)
val mask = outBufs[iMask].readFloat()        // [19248*32] coefficients
val proto = outBufs[iProto].readFloat()      // [138*138*32] prototypes

// host-side decode (priors.bin bundled as an asset):
//   box = SSD-decode(loc, priors, variances=[0.1,0.2]); per-class NMS (score 0.3, IoU 0.5);
//   per kept det: mask = sigmoid(proto @ coeff) (>0) cropped to the box.
// Full implementation: YolactSegmenter.kt in the sample app.

Python (LiteRT / ai-edge-litert)

import numpy as np
from ai_edge_litert.interpreter import Interpreter

it = Interpreter(model_path="yolact.tflite"); it.allocate_tensors()
inp, out = it.get_input_details(), it.get_output_details()
it.set_tensor(inp[0]["index"], x)          # [1,3,550,550] BGR, normalized (see above)
it.invoke()
outs = {tuple(o["shape"][1:]): it.get_tensor(o["index"])[0] for o in out}
loc  = outs[(19248, 4)]; conf = outs[(19248, 81)]
mask = outs[(19248, 32)]; proto = outs[(138, 138, 32)]

priors = np.fromfile("priors.bin", np.float32).reshape(-1, 4)
cxy = priors[:, :2] + loc[:, :2] * 0.1 * priors[:, 2:]
wh  = priors[:, 2:] * np.exp(loc[:, 2:] * 0.2)
boxes = np.concatenate([cxy - wh / 2, cxy + wh / 2], 1)      # x1y1x2y2 (0..1)
# then per-class NMS on conf, and mask_i = sigmoid(proto @ mask[i]) cropped to boxes[i]

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	138 / 138	~41 ms
TFLite benchmark_model (TfLiteGpuDelegateV2)	GPU (OpenCL)	138 / 138	130.4 ms
TFLite benchmark_model	CPU (XNNPACK, 4 threads)	—	1426.2 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 2.67x faster than the GPU (14.34 ms against 38.37 ms) and loads 13.02x faster (158 ms against 2060 ms).

backend	compiled	inference (median / min)	load
NPU (Hexagon v81)	on-device JIT	14.34 ms / 13.89 ms	158 ms
GPU (Adreno)	—	38.37 ms / 31.57 ms	2060 ms

Measured on a Samsung Galaxy S26 (Snapdragon 8 Elite Gen 5 / SM8850, Hexagon v81, Android 16) with 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 0.77, where 1.0 is the throttling threshold.

The NPU rows ran the published file unchanged. LiteRT compiled it for the Hexagon on the device at first load. That first compile took 12 s here. The load column above is the cached load every later run pays. Recipe and the runtime libraries it needs: NPU guide.

GPU wiring: GPU guide.

Raspberry Pi 5 (CPU)

Measured on a Raspberry Pi 5 Model B Rev 1.1 (8 GB, Raspberry Pi OS 64-bit) with the LiteRT benchmark_model tool from litert-cli-nightly 0.2.0.dev20260805: CPU inference (XNNPACK, 4 threads), 3 invocations per file of 10 warm-up plus 50 timed runs (the tool caps a phase at 150 s, so very slow graphs run fewer — the Runs column is the actual timed total). The latency is the median across invocations; the spread is the min–max over all timed runs. No thermal throttling occurred during these runs (vcgencmd get_throttled stayed 0x0).

File	Inference (median)	Spread (min–max)	Runs	Peak memory
yolact.tflite	1,095.6 ms	1,077.6–1,107.9 ms	150	376 MB

License

MIT (YOLACT / dbolya/yolact). COCO class taxonomy.