yjh0410 1fe1920305 fix the bug in the BCELoss used in Matcher		2 年之前
..
README.md	a41a710e5a debug	2 年之前
build.py	578dc295f4 redesign RTCDet	2 年之前
loss.py	58f47f0b20 update	2 年之前
matcher.py	1fe1920305 fix the bug in the BCELoss used in Matcher	2 年之前
rtcdet.py	f6753b0e9b train RTCDet-S	2 年之前
rtcdet_backbone.py	6752a50d89 update	2 年之前
rtcdet_basic.py	9650add52f fix a bug in YoloBottleNeck	2 年之前
rtcdet_head.py	578dc295f4 redesign RTCDet	2 年之前
rtcdet_neck.py	578dc295f4 redesign RTCDet	2 年之前
rtcdet_pafpn.py	62a5b79e6a update	2 年之前
rtcdet_pred.py	f6753b0e9b train RTCDet-S	2 年之前

RTCDet: My Empirical Study of Real-Time Convolutional Object Detectors.

Model	Scale	Batch	AP^{test 0.5:0.95}	AP^test 0.5	AP^val 0.5:0.95	AP^val 0.5	FLOPs ^(G)	Params ^(M)	Weight
RTCDet-N	640	8xb16	36.8	54.9	36.6	54.6	8.2	2.1	ckpt
RTCDet-T	640	8xb16	41.9	60.6	41.6	60.1	17.9	4.7	ckpt
RTCDet-S	640	8xb16	45.2	64.0	44.7	63.7	31.5	8.4	ckpt
RTCDet-M	640	8xb16					82.3	22.9
RTCDet-L	640	8xb16
RTCDet-X	640	8xb16

For training, we train my RTCDet series series with 300 epochs on COCO.
For data augmentation, we use the large scale jitter (LSJ), Mosaic augmentation and Mixup augmentation, following the setting of YOLOX, but we remove the rotation transformation which is used in YOLOX's strong augmentation.
For optimizer, we use AdamW with weight decay 0.05 and base per image lr 0.001 / 64.
For learning rate scheduler, we use linear decay scheduler.
Due to my limited computing resources, I can not train RTCDet-X with the setting of batch size=128.

Model	Scale	Batch	AP^val 0.5:0.95	AP^val 0.5	FLOPs ^(G)	Params ^(M)	Weight
RTCDet-P	320	8xb16	25.4	40.1	1.00	1.07	-
RTCDet-P	416	8xb16	29.4	45.8	1.68	1.07	-
RTCDet-P	512	8xb16	32.0	49.2	2.55	1.07	-
RTCDet-P	640	8xb16	33.8	51.7	4.05	1.07	ckpt