Review: FPN — Feature Pyramid Network (Object …
https://towardsdatascience.com/review-fpn-feature-pyramid-network-object-detection-262fc7482610
1. Comparisons with Baselines2. Top-Down Enrichment3. Lateral Connections4. Pyramid Representations (d) is the FPN but without the top-down pathway. With this modification, the 1×1 lateral connections followed by 3×3 convolutions are attached to the bottom-up pyramid. It simulates the effect of r...The results are just inferior compared with FPN (c). (d) is the FPN but without the top-down pathway. With this modification, the 1×1 lateral connections followed by 3×3 convolutions are attached to the bottom-up pyramid. It simulates the effect of r... The results are just inferior compared with FPN (c). It is conjectured that this is because there are large semantic gaps between different levels o… (d) is the FPN but without the top-down pathway. With this modification, the 1×1 lateral connections followed by 3×3 convolutions are attached to the bottom-up pyramid. It simulates the effect of r... The results are just inferior compared with FPN (c). It is conjectured that this is because there are large semantic gaps between different levels on the bottom-up pyramid (b), especially for very deep ResNets. A variant of (d) without sharing the parameters of the heads is also evaluated, but observed similarly degraded performance.
(d) is the FPN but without the top-down pathway. With this modification, the 1×1 lateral connections followed by 3×3 convolutions are attached to the bottom-up pyramid. It simulates the effect of r...
The results are just inferior compared with FPN (c).
It is conjectured that this is because there are large semantic gaps between different levels o…
It is conjectured that this is because there are large semantic gaps between different levels on the bottom-up pyramid (b), especially for very deep ResNets.
A variant of (d) without sharing the parameters of the heads is also evaluated, but observed similarly degraded performance.
DA: 60 PA: 5 MOZ Rank: 89