yjh0410 7ccf4060b5 train RTCDet-v1-X		2 lat temu
..
README.md	7ccf4060b5 train RTCDet-v1-X	2 lat temu
build.py	2eb02320fc release RTCDet-v1 & rename RTMDet to RTCDet	2 lat temu
loss.py	2eb02320fc release RTCDet-v1 & rename RTMDet to RTCDet	2 lat temu
matcher.py	2eb02320fc release RTCDet-v1 & rename RTMDet to RTCDet	2 lat temu
rtcdet_v1.py	2eb02320fc release RTCDet-v1 & rename RTMDet to RTCDet	2 lat temu
rtcdet_v1_backbone.py	2eb02320fc release RTCDet-v1 & rename RTMDet to RTCDet	2 lat temu
rtcdet_v1_basic.py	2eb02320fc release RTCDet-v1 & rename RTMDet to RTCDet	2 lat temu
rtcdet_v1_head.py	2eb02320fc release RTCDet-v1 & rename RTMDet to RTCDet	2 lat temu
rtcdet_v1_neck.py	2eb02320fc release RTCDet-v1 & rename RTMDet to RTCDet	2 lat temu
rtcdet_v1_pafpn.py	2eb02320fc release RTCDet-v1 & rename RTMDet to RTCDet	2 lat temu
rtcdet_v1_pred.py	2eb02320fc release RTCDet-v1 & rename RTMDet to RTCDet	2 lat temu

RTCDet-v1: My First Empirical Study of Real-Time Convolutional Object Detectors.

For training, we train my RTCDetv1 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 RTCDetv1-X with the setting of batch size=128.

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
RTCDetv1-N	640	8xb16	35.7	53.8	35.6	53.8	9.1	2.4	ckpt
RTCDetv1-T	640	8xb16	40.5	59.1	40.3	59.1	19.0	5.1	ckpt
RTCDetv1-S	640	8xb16	43.6	62.6	43.3	62.6	33.6	9.0	ckpt
RTCDetv1-M	640	8xb16	48.3	67.0	48.1	66.9	87.4	23.6	ckpt
RTCDetv1-L	640	8xb16	50.2	68.6	50.0	68.4	176.6	47.6	ckpt
RTCDetv1-X	640	8xb12