"""Render duration-weighted Gaussian gaze directly on the input image grid.""" import torch def make_image_heatmap(xy,durations,image_size=(512,512),sigma_px=16.): """xy: Nx2 normalized coordinates (x,y); durations: N nonnegative weights. Sigma is in pixels on the resized model-input image, not DINO patch units. The duration-weighted Gaussian sum is normalized by its maximum to [0,1]. No eye-tracker coordinates are inferred from a coarse cached heatmap. """ xy=torch.as_tensor(xy,dtype=torch.float32).reshape(-1,2) durations=torch.as_tensor(durations,dtype=torch.float32,device=xy.device).reshape(-1) height,width=image_size if len(xy)!=len(durations) or height<=0 or width<=0 or sigma_px<=0: raise ValueError('Invalid fixation lengths, image dimensions or Gaussian sigma.') if not torch.isfinite(xy).all() or not torch.isfinite(durations).all(): raise ValueError('Fixation coordinates and durations must be finite.') if (xy<0).any() or (xy>1).any() or (durations<0).any(): raise ValueError('Coordinates must be normalized to [0,1]; durations must be nonnegative.') # A separable Gaussian evaluates the same weighted sum without a costly # fixation-by-image stack. This also handles an empty fixation sequence. xx=torch.arange(width,device=xy.device)[None,:] yy=torch.arange(height,device=xy.device)[None,:] gx=torch.exp(-(xx-xy[:,0:1]*(width-1)).square()/(2*sigma_px**2)) gy=torch.exp(-(yy-xy[:,1:2]*(height-1)).square()/(2*sigma_px**2)) heatmap=(gy.transpose(0,1)*durations[None,:])@gx return (heatmap/heatmap.max().clamp_min(1e-12)).unsqueeze(0)