EV74 visual frontend Nodes

Neat exposes the EV74 visual-frontend graphs as normal Graph Nodes. Use the public Node factories and option structs; do not call processcvu, ConfigManager, or dispatcher APIs directly from application code.

Node factory	Graph name	Graph ID	Purpose
nodes::FeatureHistogram / pyneat.nodes.feature_histogram	feature_histogram	235	Grayscale image histogram
nodes::GriderFast / pyneat.nodes.grider_fast	grider_fast	236	Grid-distributed FAST features
nodes::TrackDescriptor / pyneat.nodes.track_descriptor	track_descriptor	237	FAST features plus descriptors
nodes::TrackKLT / pyneat.nodes.track_klt	track_klt	238	Pyramidal KLT tracking, optionally with detected replacement features

Graph IDs are useful for diagnostics and firmware/package parity checks. They are not required in application code.

Tensor contract

All tensors use logical batch shapes. If batch_size == B, a grayscale image is [B,H,W], not [B*H,W]. The runtime handles any EV74 transport packing internally.

Node	Inputs	Public outputs
FeatureHistogram	input_image: UInt8 [B,H,W]	output_hist: Int32 [B,256]
GriderFast	input_image: UInt8 [B,H,W]	output_features: Int32 [B,1 + max_features*3]
TrackDescriptor	input_image: UInt8 [B,H,W]	output_features: Int32 [B,1 + max_features*3]; output_descriptors: Int32 [B,max_features,8]
TrackKLT	prev_image: UInt8 [B,H,W]; cur_image: UInt8 [B,H,W]; input_points: Int32 [B,num_points,2]	output_points: Float32 [B,num_points,2]; output_status: Int32 [B,num_points,1]; plus output_features: Int32 [B,1 + max_features*3] only when detect_new_features != 0

Feature-list tensors use this per-batch layout:

[count, x0, y0, score0, x1, y1, score1, ...]

The descriptor graph currently requires descriptor_words == 8. Changing that is an EV74 ABI change and is rejected before dispatch.

C++ quick start

#include <neat.h>

#include <cstdint>
#include <vector>

using namespace simaai::neat;

Tensor make_gray_batch(int width, int height, int batch) {
  std::vector<std::uint8_t> pixels(static_cast<std::size_t>(width) * height * batch);
  // Fill pixels in batch-major order: b*height*width + y*width + x.
  auto tensor = Tensor::from_vector(pixels, {batch, height, width}, TensorMemory::EV74);
  tensor.layout = TensorLayout::HW;
  tensor.axis_semantics = {TensorAxisSemantic::N, TensorAxisSemantic::H, TensorAxisSemantic::W};
  tensor.route.name = "input_image";
  tensor.route.segment_name = "input_image";
  return tensor;
}

int main() {
  constexpr int width = 320;
  constexpr int height = 240;
  constexpr int batch = 2;

  Graph graph;

  InputOptions input;
  input.payload_type = PayloadType::Tensor;
  input.format = FormatTag::UINT8;
  input.width = width;
  input.height = height;
  input.depth = 1;
  input.max_width = width;
  input.max_height = height * batch; // transport capacity; public tensor remains [B,H,W]
  input.max_depth = 1;
  input.memory_policy = InputMemoryPolicy::Ev74;
  input.buffer_name = "input_image";

  graph.add(nodes::Input(input));

  GriderFastOptions fast;
  fast.width = width;
  fast.height = height;
  fast.batch_size = batch;
  fast.max_features = 64;
  fast.threshold = 30;
  graph.add(nodes::GriderFast(fast));

  graph.add(nodes::Output());

  RunOptions run_opt;
  run_opt.output_memory = OutputMemory::Owned;

  Tensor image = make_gray_batch(width, height, batch);
  Run run = graph.build({image}, run_opt);
  TensorList outputs = run.run({image}, /*timeout_ms=*/30000);
  run.close();
}

KLT with three inputs

TrackKLT consumes a tensor-set: previous image, current image, and input points. Name the routes to match the option fields.

TrackKLTOptions klt;
klt.width = 320;
klt.height = 240;
klt.batch_size = 2;
klt.num_points = 32;
klt.max_features = 64;
klt.detect_new_features = 1; // publish output_features as the third output

graph.add(nodes::TrackKLT(klt));

Expected public outputs when detect_new_features == 1:

output_points   Float32 [2,32,2]
output_status   Int32   [2,32,1]
output_features Int32   [2,193]

When detect_new_features == 0, Neat publishes only output_points and output_status; the EV-visible features buffer remains an internal runtime allocation.

Python surface

The Python API mirrors the C++ options/factory style and is intentionally layer-like: create an options object, set public configuration, and add the Node to a Graph.

import numpy as np
import pyneat

width, height, batch = 320, 240, 2

opt = pyneat.GriderFastOptions()
opt.width = width
opt.height = height
opt.batch_size = batch
opt.max_features = 64
print(opt.summary())

graph = pyneat.Graph()
input_opt = pyneat.InputOptions()
input_opt.payload_type = pyneat.PayloadType.Tensor
input_opt.format = "UINT8"
input_opt.width = width
input_opt.height = height
input_opt.max_width = width
input_opt.max_height = height * batch
input_opt.memory_policy = pyneat.InputMemoryPolicy.Ev74
input_opt.buffer_name = "input_image"

graph.add(pyneat.nodes.input(input_opt))
graph.add(pyneat.nodes.grider_fast(opt))
graph.add(pyneat.nodes.output())

image_np = np.zeros((batch, height, width), dtype=np.uint8)
image = pyneat.Tensor.from_numpy(image_np, memory="ev74")
image.layout = pyneat.TensorLayout.HW
# If setting route metadata from Python in a custom app, keep it aligned with
# the option names used above.

Safety checks

These Nodes validate graph envelopes before EV dispatch. They reject:

• non-positive dimensions or counts;
• unsupported batch sizes;
• thresholds outside [0,255];
• duplicate/empty tensor names;
• TrackDescriptorOptions.descriptor_words != 8;
• invalid KLT window, level, and detect-mode values;
• undersized runtime tensors during pre-dispatch negotiation.

This matters because illegal buffers can wedge EV74. Keep validation failures as host-side errors and do not bypass the Node contract path.

Fast validation command

The fast customer-style DevKit gate is:

ctest --test-dir /workspace/core_graph_changes/build/tests \
  -R visual_frontend_ --output-on-failure

It runs:

• all four visual graphs with 320x240, batch_size=2, detect_new_features=1;
• a targeted KLT no-detect ABI check;
• a negative pre-dispatch guard check that confirms illegal batch input is rejected before EV74.