Skip to main content

Architecture Overview

SpikeSEG is a spiking encoder-decoder network with an instance-segmentation stage (HULK-SMASH) built on top of the encoder.

Full Pipeline

Layer Summary (IGARSS 2023)

LayerTypeKernelChannelsThresholdLeakNeuron
Conv1Spiking Conv2d5x5410.09.0 (90%)LIF subtractive
Pool1Max Pool2x2--------
Conv2Spiking Conv2d5x53610.01.0 (10%)LIF subtractive
Pool2Max Pool2x2--------
Conv3Spiking Conv2d7x7nclassesn_\text{classes}10.00.0IF

Weight Initialization

All convolutional weights are drawn from N(0.8,0.01)\mathcal{N}(0.8, 0.01) and clipped to [0,1][0, 1]. Starting near 1 accelerates STDP convergence.

Two Output Paths

  1. Saliency map (decoder path): all classification spikes are decoded together into a single pixel-level heat map.
  2. Instance masks (HULK-SMASH path): each classification spike is decoded separately, producing one mask per instance, then grouped by SMASH similarity.

See Encoder, Decoder, HULK-SMASH.