On-device crowd counting running fully on the LiteRT CompiledModel GPU delegate
(no CPU fallback). DM-Count (NeurIPS 2020)
regresses a person density map whose sum is the crowd size — it counts hundreds of
people where detector-based counting saturates.
Architecture: VGG19 backbone + conv regression head — pure CNN.
Output:[1, 1, 64, 64] non-negative density map — sum(map) = estimated person count; normalize per frame for a heatmap overlay.
GPU conversion
DM-Count is a pure CNN (VGG19 + conv head). It converts fully GPU-compatible (30/30 nodes
on the delegate, 1 partition; Pixel 8a corr 0.9998–1.0 and count within 0.4% of PyTorch on
real crowd images, ~79 ms/frame) with one exact rewrite: the mid-graph
F.upsample_bilinear (align_corners=True RESIZE_BILINEAR, banned on the delegate) is a
linear operator, re-authored as two constant-matrix multiplies — with the constant on the
RHS (lowers to FULLY_CONNECTED; the delegate rejects BATCH_MATMUL with a constant
LHS). Desktop corr vs PyTorch is 1.000000 with an identical count.
Minimal usage
Kotlin (Android, LiteRT CompiledModel GPU)
kotlin
1val options = CompiledModel.Options(Accelerator.GPU)2val model = CompiledModel.create(context.assets,"dmcount.tflite", options,null)3val inBufs = model.createInputBuffers()4val outBufs = model.createOutputBuffers()56inBufs[0].writeFloat(inputNCHW)// [1,3,512,512] RGB, ImageNet-norm7model.run(inBufs, outBufs)8val density = outBufs[0].readFloat()// [64*64] density map9val count = density.sum()// estimated number of people
Converted with litert-torch (build_dmcount.py): loads the MIT DM-Count (UCF-QNRF)
weights and exports the raw density map. The UCF-QNRF checkpoint generalizes best across
scenes; the upstream repo also bundles an NWPU-Crowd variant.
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
30 / 30
~79 ms
TFLite benchmark_model (TfLiteGpuDelegateV2)
GPU (OpenCL)
30 / 30
98.2 ms
TFLite benchmark_model
CPU (XNNPACK, 4 threads)
—
3185.7 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 2.90x faster than the GPU (9.54 ms against 27.64 ms) and loads 6.79x faster (139 ms against 943 ms).
backend
inference (median / min)
load
NPU (Hexagon v81)
9.54 ms / 9.16 ms
139 ms
GPU (Adreno)
27.64 ms / 27.38 ms
943 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 0.70-0.71, 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.
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
dmcount.tflite
1,923.8 ms
1,858.0–2,054.7 ms
150
367 MB
License
MIT (DM-Count / cvlab-stonybrook). Trained on UCF-QNRF.
