ramdrop / autoplace Goto Github PK

Thanks for your codes @ramdrop . I noticed some manner in common.py, when doing the knn search. The corresponding code is as followings:

def cal_recall_feat(gt, qFeat, dbFeat):
    # --------------------------------- use faiss to do NN search -------------------------------- #
    # faiss_index = faiss.IndexFlatL2(qFeat.shape[1])
    # faiss_index.add(dbFeat)
    # dists, predictions = faiss_index.search(qFeat, 20)   # the results is sorted
    # -------------------------------- use sklearn to do NN search ------------------------------- #
    knn = NearestNeighbors(n_jobs=1)
    knn.fit(dbFeat)
    dists, predictions = knn.kneighbors(qFeat, 20)
    recalls = cal_recall_pred(gt, predictions)
    return recalls, dists, predictions

Can you share me some advice about how to use them? Thanks!

Hello, I installed conda ,python environment and downloaded the dataset, but it is running/ gene_ woDTR. SH times wrong.
AssertionError: Database version not found: v1.0-trainval

Hi @ramdrop , thanks for your amazing works and the release of the code. I find some positions that quite confuse me, and looking forward to your answer.

1. here, why the number of search is len(dbFeat), in my opinion, the n_values should be used?

    faiss_index = faiss.IndexFlatL2(opt.output_dim)
    faiss_index.add(dbFeat)
    # dists, predictions = faiss_index.search(qFeat, max(n_values))   # the results is sorted
    dists, predictions = faiss_index.search(qFeat, len(dbFeat))

2. I wonder why the radius dramatically change the PR performance during getting the GT?

def getPositives(self):
        # positives for evaluation are those within trivial threshold range
        # fit NN to find them, search by radius
        if self.positives is None:
            knn = NearestNeighbors(n_jobs=-1)
            knn.fit(self.dbStruct.utmDb)
            self.distances, self.positives = knn.radius_neighbors(self.dbStruct.utmQ, radius=self.dbStruct.nonTrivPosDistSqThr**0.5) 
        return self.positives

Hello, thank you for sharing your code. But I don't quite understand why you didn't use all samples to generate the database?

Hello, I have some questions about the ". Pickle" file in the folder.
Question 1: Figure 1 is about the "features_test. Pickle" produced by "evaluate" after "se_dpr" training. I will open the line beginning with "gt": "1150、1151、2197、1148、1149、2196、2195、2194" respectively. The pictures under the corresponding "dateset / img" are shown in Figure 2. From the point cloud information shown in the figure, can it be understood that the upper row belongs to one scene and the lower row belongs to another scene? So why put them in an array in the ". Pickle" file?And they are not in positive order. For example, "2197" appears between "1151"and "1151".
Question 2: Figure 3 shows the data information of "se_te_results. Pickle". In the array of "preds", "1150,1277,4313" appear in "datebase" and "6113" appear in "test", but why do the numbers "37 "and "38 "not appear in "- test", "val", "train" and "- datebase"? What information do these numbers represent? Do they refer to the verified data set? Or other? I don't quite understand. And why do the middle of the data pass "..." Separated by ellipsis?

Dear Sir,

The paper experiments used radar data from nuScenes.
What is your opinion on using your method for 4D radar point clouds?
These point clouds have height info and often a 120 deg Horizontal FOV and 30 deg vertical FOV.

Thanks,

Hi author，
In formula (1), alpha is the heading Angle. Could you please tell me how this Angle is defined and which coordinate system is used as the reference? The sensor coordinate system, or the coordinate system at the rear of the vehicle.
Tks for your attention~