fortunechen / reproduced_object_detection

object detection, mmdetection

To reproduce some important object detection experiment.

To do list:

Two stages:

backbone	name	FPN	cascade	tricks	AP50:90
res50	Faster RCNN	✔			35.5
res50	Faster RCNN	✔		dconv_c3-c5	40.1
res50	Faster RCNN	✔	✔		40.5
res101	Faster RCNN	✔			38.6
res101	Faster RCNN	✔	✔		41.5

One stage:

backbone	name	tricks	AP50:90
res50	RetinaNet		35.6
res101	RetinaNet		37.6
VGG16	SSD300		25.4
VGG16	SSD512		29.5

proposal

backbone	name	tricks	AR100	AR300	AR1000(mmdetection)
res50	RPN		42.82	51.44	57.25(57.1)
res101	RPN		0.4545	0.5337	58.66(58.6)

---

1. Faster RCNN(FPN)

backbone	our AP 0.50:0.95(mmdetection, paper)	training time
ResNet50	35.5(36.4, 33.9)	26h (4 titanxp)
ResNet101	38.6 (38.6, 36.2)	64h (2 titanxp)

ResNet50

ResNet101

---

2. Cascade-RCNN(FPN)

backbone	AP 0.50:0.95(mmdetection, paper)	training time
ResNet50	40.5(40.3,40.3)	34h (8 titanxp , nondistributed)
ResNet101	41.5(42.1, 42.7)	72h (2 titanxp)

ResNet50

ResNet101

---

deformable conv or pooling

backbone	our AP 0.50:0.95(mmdetection, paper)	training time
ResNet50	40.1 ()	68h(2 tianxp, distributed)

---

3. RetinaNet

backbone	AP 0.50:0.95(mmdetection, paper)	training time
ResNet50	35.6(35.8, 35.7)	25h(4 titanxp)
ResNet101	37.6(37.7, 37.8)	33h(4 titanxp)

ResNet50

ResNet101

---

4. SSD

image size	AP 0.50:0.95(mmdetection, paper)	training time
300	25.4(25.7, 23.2)	71h(4 titanxp, 24epoch)
512	29.5(29.3, 26.8)	73h(4 titanxp, 24epoch)

SSD300

SSD512

---

if not noted, the default setting is 12 epoch with distributed training.