When output the final depth prediction, it seems you go through a conv layer and output without going through an activation function. This makes the evaluation in the log10 space get NaN or Inf, because some of the output predictions have a negative value even though there shouldn't be negative depth value.
When output the final depth prediction, it seems you go through a conv layer and output without going through an activation function. This makes the evaluation in the log10 space get NaN or Inf, because some of the output predictions have a negative value even though there shouldn't be negative depth value.