Recursively Refined R-CNN: Self RoI Re-Balancing on Instance Segmentation

Installation

See MMDetection and MMCV official pages to know the installation procedure for mmdetection and mmcv libraries.

Our code has been designed with the following version of the libraries:

• MMDetection - commit eb66e0d0
• MMCV - commit 1830347

Note: In order to run R³-CNN training, it is required to install our mmcv library version (not the official one!).

Models

Most important files.

Name	Description
FCC BBox Head	It contains our brand new R3-CNN bbox head with FCC Lighter v1 and v2, corresponding to Fig. 4(b, c).
FCC BBox Head	It contains our brand new R3-CNN bbox head with FCC Advanced v1 and v2, corresponding to Fig. 4(d, e).
Mask Head	It contains R3-CNN mask head.
FCC Mask IoU Head	It contains our brand new Mask IoU segmentation head for R3-CNN model with FCC Lighter v1 and v2, corresponding to Fig. 4(b, c).
FCC Mask IoU Head	It contains our brand new Mask IoU segmentation head for R3-CNN model with FCC Advanced v1 and v2, corresponding to Fig. 4(d, e).
R3-CNN RoI Head	It contains R3-CNN RoI Head whose task is to manage bbox, segmentation and Mask IoU heads.

Experiments

The following configuration files have been used to run each experiment:

Table 2

Row	Model
1	Mask (1x)
2	Mask (3x)
3	HTC
:--:	:----:
4	R3-CNN (naive)
5	R3-CNN (deeper)

Table 3

Row	Model	Lt	H	Alt.
1	HTC	3	3	abc
:--:	:--:	:----:	:--:	:--:
2	R3-CNN	2	2	ab
3	R3-CNN	3	2	abb
4	R3-CNN	3	2	aab
5	R3-CNN	3	2	aba
6	R3-CNN	4	2	aabb
7	R3-CNN	4	2	abab
8	R3-CNN	5	2	aabbb

Table 4

Row	Model	Lt	H	L2C	NLb	NLa
1	HTC	3	3
:--:	:----:	:--:	:--:	:----:	:--:	:--:
2	R3-CNN	3	1
3	R3-CNN	3	1	7x7
4	R3-CNN	3	1	7x3 -> 3x7
5	R3-CNN	3	1	7x7	v
6	R3-CNN	3	1	7x3 -> 3x7	v
7	R3-CNN	3	1	7x7	v	v
:--:	:----:	:--:	:--:	:----:	:--:	:--:
8	R3-CNN	4	2
9	R3-CNN	4	2	7x7
10	R3-CNN	4	2	7x3 -> 3x7
11	R3-CNN	4	2	7x7	v
12	R3-CNN	4	2	7x3 -> 3x7	v

Table 5

| Row | Model | Lt | H | M_IoU | L2C | NL_b | NL_a | :--: | :----: | :--: | :--: | :----: | :--: | :--: | | 1 | HTC | 3 | 3 | | | | | | :--: | :----: | :--: | :--: | :----: | :--: | :--: | :--: | | 2 | R³-CNN | 3 | 1 | | | | | | 3 | R³-CNN | 3 | 1 | v | | | | | 4 | R³-CNN | 3 | 1 | v | 7x7 | | | | 5 | R³-CNN | 3 | 1 | v | 7x3 -> 3x7 | | | | 6 | R³-CNN | 3 | 1 | v | 7x7 | v | | | 7 | R³-CNN | 3 | 1 | v | 7x7 | v | v | | :--: | :----: | :--: | :--: | :----: | :--: | :--: | :--: | | 8 | R³-CNN | 4 | 2 | | | | | | 9 | R³-CNN | 4 | 2 | v | | | | | 10 | R³-CNN | 4 | 2 | v | 7x7 | | | | 11 | R³-CNN | 4 | 2 | v | 7x7 | v | | | 12 | R³-CNN | 4 | 2 | v | 7x7 | v | v |

Table 6

Model
baseline
conv 3x3
conv 5x5
conv 7x7
conv 7x3 -> 3x7
Non-local 1x1
Non-local 3x3

Table 7

Model
baseline
conv 3x3
conv 5x5
conv 7x7
conv 7x3 -> 3x7
Non-local 1x1
Non-local 3x3

Table 8

Model
baseline
GRoIE

Table 9

Row	Model	Backbone
1	Mask	r50-FPN
2	HTC	r50-FPN
3	SBR-CNN	r50-FPN
:--:	:----:	:------:
6	GC-Net	r50-FPN
7	HTC + GC-Net	r50-FPN
8	SBR-CNN + GC-Net	r50-FPN
:--:	:----:	:------:
9	DCN	r50-FPN
10	HTC + DCN	r50-FPN
11	SBR-CNN + DCN	r50-FPN