aipixel / GPS-Gaussian

[CVPR 2024 Highlight] The official repo for “GPS-Gaussian: Generalizable Pixel-wise 3D Gaussian Splatting for Real-time Human Novel View Synthesis”

Home Page:https://shunyuanzheng.github.io/GPS-Gaussian

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

关于渲染实时性

yqxuxuxu opened this issue · comments

作者您好,我使用以下代码对渲染实时性进行了测试,但是结果只有4.fps,远没有论文列出的25fps这么高,不知道是我的操作有问题还是因为卡不一样呢?

        fps = []
        for idx in tqdm(range(total_frames)):
            item = self.dataset.get_test_item(idx, source_id=view_select)
            data = self.fetch_data(item)

            data = get_novel_calib(data, self.cfg.dataset, ratio=ratio, intr_key='intr_ori', extr_key='extr_ori')
            start_time = time.time() 
            with torch.no_grad():
                data, _, _ = self.model(data, is_train=False) 
                data = pts2render(data, bg_color=self.cfg.dataset.bg_color)
            end_time = time.time()  
            synthesis_time = end_time - start_time
            fps.append(1/synthesis_time)
            render_novel = self.tensor2np(data['novel_view']['img_pred'])
            
            cv2.imwrite(test_path  + '/%s_novel.jpg' % (data['name']), render_novel)

        print(sum(fps)/ len(fps))

python版本的差不多在5fps左右,实时系统需要把网络部分全部转成TensorRT,跟FloRen中类似。

python版本的差不多在5fps左右,实时系统需要把网络部分全部转成TensorRT,跟FloRen中类似。

好的,非常感谢。