"segment-anything" is slow

[GitChanyoung]

Hello, I'm grateful for your research. I tried segment-anything with the code you shared. I though it would be fast, but it shows a very slow speed with an average of 13000ms. Can you tell me the reason why?

from efficient_sam.build_efficient_sam import build_efficient_sam_vits
import zipfile

with zipfile.ZipFile("weights/efficient_sam_vits.pt.zip", 'r') as zip_ref:
    zip_ref.extractall("weights")
efficient_sam_vits_model = build_efficient_sam_vits()
efficient_sam_vits_model.eval()

import os
image_path = "dataset/data1"
image_list = os.listdir(image_path)

import time
for image_name in image_list:
    image_path1 = os.path.join(image_path, image_name)
    st = time.time()
    mask_efficient_sam_vits = run_everything_ours(image_path1, efficient_sam_vits_model)
    et = time.time()
    print('inference time:', (et - st)*1000)

Output:

inference time: 17762.072563171387
inference time: 13303.216695785522
inference time: 13251.076221466064
inference time: 12843.31202507019

image_size : 640x640 GPU : RTX A6000

[yformer]

@GitChanyoung, thanks for your interest! We will take a look for segment everything.

[silinsi]

I experienced similar thing, it seems that code provided in [EfficientSAM_segment_everything_example.ipynb] is not adapted to gpu , i tried to move the network to gpu but this does not help a lot( probably because my modification not complete ), saving around ten seconds but it still costs around five to ten seconds. Moreover, to the same image, it runs a little faster after first run.

[klightz]

Hi, thanks for your interest, the segment-anything notebook example is work-in-progress and currently it only supports CPU version (you can find that tensors are forced to cpu in the function) with a naive & straight forward algorithm. We will adapt a quicker version in the future.

[sulaimanvesal]

When running on a CPU the Tiny EfficienetSAM, the performance of this model is significantly slower than that of FastSAM-S. After conducting a thorough comparison between FastSAM and EfficientSAM.

The following result is running the code on Google Collab:

Inference using:  efficientsam_ti_cpu.jit
Input size: torch.Size([3, 512, 1024])
Preprocess Time: 79.8783 ms
Inference Time: 6939.1549 ms

[sipie800]

same here, more than 30s on rtx3090, not efficient at all

[liutongkun]

same here, more than 30s on rtx3090, not efficient at all

on my 3090ti，i7 12700， it's 30-40ms per image. The problem is that the speed will be slower when running the first inference, I'm not sure why