Method __getitem__

datasets/provider_vos.py:100–185

(self, index)

        return len(self.ret_frames)

    def __getitem__(self, index):
        set_idx, image_path, timestamp, frame_idx = self.ret_frames[index]
        
        if not osp.exists(image_path):
            raise FileNotFoundError(f"Image not found: {image_path}")
        
        main_image = Image.open(image_path).convert("RGB")
        depth_path = self.get_predicted_depth_path(image_path)
        depth_image = Image.open(depth_path)
        main_image = self.transform(main_image)
        depth_image = self.depth_transform(depth_image)

        if self.training:
            mask_path = image_path.replace(self.images_root, self.mask_root).replace(".jpg", ".png")
            if not osp.exists(mask_path):
                raise FileNotFoundError(f"Mask not found: {mask_path}")
            mask_image = Image.open(mask_path)
            mask_image = self.transform_mask(mask_image).unsqueeze(0)
        else:
            mask_image = torch.ones_like(depth_image)
        
        frames = [main_image]
        depths = [depth_image]
        masks = [mask_image]
        timestamps_list = [0.0]
        
        seq_length = self.set_len[set_idx]
        current_idx = frame_idx
        
        if self.opt.output_frames > 1 and seq_length > 1:
            if not self.shuffle:
                # pick frame with fixed interval
                if current_idx + self.nearby_range >= seq_length:
                    interval = (seq_length - current_idx) // (self.opt.output_frames - 1)
                else:
                    interval = self.nearby_range // (self.opt.output_frames - 1)
                assert interval > 0, "Interval must be greater than 0"
                offsets = [i * interval for i in range(1, self.opt.output_frames)]
            else:
                if current_idx + self.nearby_range >= seq_length:
                    offsets = random.sample(range(1, seq_length - current_idx), self.opt.output_frames - 1)
                else:
                    offsets = random.sample(range(1, self.nearby_range + 1), self.opt.output_frames - 1)

            offsets.sort() 
            
            for offset in offsets:
                pair_idx = current_idx + offset
                img_file_pair, ts_pair = self.frame_infos[set_idx][pair_idx]
                if not osp.exists(img_file_pair):
                    print(f"Warning: Missing file {img_file_pair}")
                    continue
                pair_image = Image.open(img_file_pair).convert("RGB")
                pair_depth_path = self.get_predicted_depth_path(img_file_pair)
                pair_depth_image = Image.open(pair_depth_path)
                pair_image = self.transform(pair_image)
                pair_depth_image = self.depth_transform(pair_depth_image)