数据集

[Velpro-collab]

你好，请问一下，MITML这个方法有没有使用auxiliary set.

[dyhBUPT]

Hi, in our experiments, the auxiliary set is not used for all other SOTA methods.

[Velpro-collab]

Hi, in our experiments, the auxiliary set is not used for all other SOTA methods. 还有一个问题query和gallery是不是包含了可见和红外样本的，

[Velpro-collab]

其他SOTA方法使用的训练测试方法，是不是和开源的代码一样，还是使用SOTA的训练方法

[dyhBUPT]

Hi, in our experiments, the auxiliary set is not used for all other SOTA methods. 还有一个问题query和gallery是不是包含了可见和红外样本的，

I don't understand your question well. Maybe you can refer to our codes for details.

[dyhBUPT]

其他SOTA方法使用的训练测试方法，是不是和开源的代码一样，还是使用SOTA的训练方法

We reimplement all SOTA methods using their own training codes.

[Velpro-collab]

作者，您好，你们制作的这个数据集对于视频可见红外行人重识别的研究很有价值，MITML这篇论文是我们同实验室师兄的作品，我们也想在你们的数据集上测试一下，以便和我们后续的工作进行对比，但是我们复现效果特别差，能不能开源一下你们在MITML这个方法的代码，如果可以将不甚感激，我的邮箱：211861921@qq.com

[dyhBUPT]

Thanks for your interest in our work, MITML is an excellent work, which greatly inspires our work. I'm sorry but I didn't clean up the codes for these reimplemented SOTA methods. You can try to convert the BUPTCampus to 'HITSZ-VCM' style first, then you can conveniently train MITML on BUPTCampus with minor modifications. Best wishes.

[Velpro-collab]

# distance
if opt.distance == 'cosine':
    distance = 1 - query_feats @ gallery_feats.T
else:
    distance = euclidean_dist(query_feats, gallery_feats)

CMC, MAP = [], []

# evaluate (intra/inter-modality)
for q_modal in (0, 1):
    for g_modal in (0, 1):
        q_mask = query_modals == q_modal
        g_mask = gallery_modals == g_modal
        tmp_distance = distance[q_mask, :][:, g_mask]
        tmp_qid = query_pids[q_mask]
        tmp_gid = gallery_pids[g_mask]
        tmp_cmc, tmp_ap = evaluate(tmp_distance, tmp_qid, tmp_gid, opt)
        CMC.append(tmp_cmc * 100)
        MAP.append(tmp_ap * 100)
        if show:
            print_metrics(
                tmp_cmc, tmp_ap,
                prefix='{:<3}->{:<3}:  '.format(MODALITY_[q_modal], MODALITY_[g_modal])
            )

# evaluate (omni-modality)
cmc, ap = evaluate(distance, query_pids, gallery_pids, opt)
CMC.append(cmc * 100)
MAP.append(ap * 100)

作者你好，这段代码里为什么cmc和ap要乘以100？

[dyhBUPT]

It's used to scale the score from [0, 1] to [0, 100]

[Velpro-collab]

Thanks for your interest in our work, MITML is an excellent work, which greatly inspires our work. I'm sorry but I didn't clean up the codes for these reimplemented SOTA methods. You can try to convert the BUPTCampus to 'HITSZ-VCM' style first, then you can conveniently train MITML on BUPTCampus with minor modifications. Best wishes.

作者，你好，我的方法是在你们开源代码的基础上仅仅将 BUPTCampus 的dataloader返回的数据形式改成 'HITSZ-VCM'的形式，但模型的训练结果是不理想，请问一下，对于 BUPTCampus 你们做了哪些数据预处理？

[dyhBUPT]

Hi, no special data processing methods are used. Please refer to these lines for data processing: https://github.com/dyhBUPT/BUPTCampus/blob/c29bb4879c8bd958ac0cd924c426e3057981a4a4/utils.py#L48-L64