My dateset map is much lower than yolov5

[tusik]

Hi, I'm using same dateset for yolov5 and yolox. Dataset include 120 cell images, 100 train and val, 20 test. Object number per image is between 50-800. And this is my exp file for yolox

class Exp(MyExp):
    def __init__(self):
        super(Exp, self).__init__()
        self.num_classes = 3
        self.depth = 1.0
        self.width = 1.0
        self.warmup_epochs = 1
        self.mosaic_prob = 1.0
        self.mixup_prob = 1.0
        self.hsv_prob = 1.0
        self.flip_prob = 0.5
        self.random_size = (28, 32) 
        self.input_size = (1280,1280)
        self.test_size = (1280,1280)
        self.exp_name = os.path.split(os.path.realpath(__file__))[1].split(".")[0]
        self.save_history_ckpt = False
        self.max_epoch = 1000
        self.eval_interval = 1
        self.basic_lr_per_img = 0.01/64.0

...
   max_labels=2000
...

Max map.5-.95 is only 0.18 yolov5-l is close to 0.6, yolov5x6-l can get to 0.7. Can you give me some solotion to imporve map?

[jaideep11061982]

@tusik did it get resolved ?

[laborer123]

I had the same problem, max map.5-.95 and map.5 vary greatly, training time also too slow, but I used yolox_s

[eladmeir]

I am getting the same problem here.... Converted a YOLOv5 dataset (with ~95% accuracy, with my measurement) using globox to COCO, and got only ~80% with YOLOx-S (which is still much bigger than the YOLOv5's size that I was using for comparison)