Segmenting objects with indistinct edges, with application to aerial imagery of vegetation

James, Katherine; Bradshaw, Karen

doi:10.1145/3351108.3351124

Cited by 2 publications

(2 citation statements)

References 8 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…A possible way of reducing this effect is to make use of a weighted BCE (WBCE) loss, in which each class’ contribution is weighted by the inverse frequency of its occurrence. This is seen in Equation .

L_{WBCE} = ‐ \frac{1}{N} {false\sum}_{n = 1}^{N} \sum_{c = 1}^{C} ω_{c} g_{italicnc} \log false(p_{nc} false) .

In the case of our binary problem,

ω_{c} = \{\begin{array}{l} 0.08 if background class \\ 0.92 if target class \end{array}) .

Furthermore, in the case where objects have indistinct edges, the use of a pre‐computed weight map to reduce the contribution of edge pixels with low annotation confidence can be used (James & Bradshaw, 2019). This is defined as in Equation .

L = ‐ \frac{1}{N} {false\sum}_{n = 1}^{N} \sum_{c = 1}^{C} ω_{map} g_{italicnc} \log false(p_{nc} false),

where ω map is defined with two additional hyper‐parameters α and β as,

ω_{map} = \{\begin{matrix} left \\ left α if pixel falls within edge of thickness t \\ left β otherwise \end{matrix} .)

The inverse class frequency weighted variant of this loss function was used, defined in Equation , where ω c has the same values as defined for WBCE.…”

Section: Methodsmentioning

confidence: 99%

See 1 more Smart Citation

Detecting plant species in the field with deep learning and drone technology

James

Bradshaw

2020

Methods Ecol Evol

Self Cite

View full text Add to dashboard Cite

show abstract

L_{WBCE} = ‐ \frac{1}{N} {false\sum}_{n = 1}^{N} \sum_{c = 1}^{C} ω_{c} g_{italicnc} \log false(p_{nc} false) .

In the case of our binary problem,

ω_{c} = \{\begin{array}{l} 0.08 if background class \\ 0.92 if target class \end{array}) .

L = ‐ \frac{1}{N} {false\sum}_{n = 1}^{N} \sum_{c = 1}^{C} ω_{map} g_{italicnc} \log false(p_{nc} false),

where ω map is defined with two additional hyper‐parameters α and β as,

ω_{map} = \{\begin{matrix} left \\ left α if pixel falls within edge of thickness t \\ left β otherwise \end{matrix} .)

The inverse class frequency weighted variant of this loss function was used, defined in Equation , where ω c has the same values as defined for WBCE.…”

Section: Methodsmentioning

confidence: 99%

“…In the case of our binary problem, Furthermore, in the case where objects have indistinct edges, the use of a pre-computed weight map to reduce the contribution of edge pixels with low annotation confidence can be used (James & Bradshaw, 2019). This is defined as in Equation 3.…”

Section: Loss Functionsmentioning

confidence: 99%