LMBN feature selection effect on MOT17 metrics

LMBN feature selection effects on mot metrics

        g_hat     = embs[:, 0:512]        # ĝ → post-BNNeck global avg pooled feature
        gdrop_hat = embs[:, 512:1024]     # ĝdrop → post-BNNeck dropblock + max pooled feature
        p1_hat    = embs[:, 1024:1536]    # p̂1 (upper body)
        p2_hat    = embs[:, 1536:2048]    # p̂2 (lower body)
        pg_hat    = embs[:, 2048:2560]    # p̂g (global part branch)
        c1_hat    = embs[:, 2560:3072]    # ĉ1 (channel partition 1)
        c2_hat    = embs[:, 3072:3584]    # ĉ2 (channel partition 2)
        embs = np.concatenate([gdrop_hat], axis=1)
        
        # g_hat, gdrop_hat, p1_hat, p2_hat, pg_hat, c1_hat, c2_hat
        # {"HOTA": 68.888, "AssA": 71.15, "AssRe": 76.626, "MOTA": 78.232, "IDSW": 145, "IDF1": 81.331, "IDs": 371}
        
        # g_hat, p1_hat, p2_hat, pg_hat
        # {"HOTA": 68.832, "AssA": 70.926, "AssRe": 76.439, "MOTA": 78.258, "IDSW": 146, "IDF1": 81.349, "IDs": 373}
        
        # gdrop_hat, p2_hat, p1_hat, pg_hat
        # {"HOTA": 68.686, "AssA": 70.6, "AssRe": 76.169, "MOTA": 78.31, "IDSW": 146, "IDF1": 81.148, "IDs": 372}
        
        # p1_hat, p2_hat, pg_hat
        # {"HOTA": 68.837, "AssA": 70.937, "AssRe": 76.454, "MOTA": 78.259, "IDSW": 145, "IDF1": 81.358, "IDs": 372}
        
        # gdrop_hat, p2_hat, pg_hat
        # {"HOTA": 68.884, "AssA": 71.152, "AssRe": 76.627, "MOTA": 78.222, "IDSW": 147, "IDF1": 81.322, "IDs": 371}
        
        # p1_hat, p2_hat
        # {"HOTA": 68.859, "AssA": 71.133, "AssRe": 76.604, "MOTA": 78.165, "IDSW": 149, "IDF1": 81.284, "IDs": 373}
        
        # g_hat
        # {"HOTA": 68.789, "AssA": 70.884, "AssRe": 76.361, "MOTA": 78.321, "IDSW": 145, "IDF1": 81.198, "IDs": 372}
        
        # gdrop_hat
        # {"HOTA": 68.751, "AssA": 70.918, "AssRe": 76.406, "MOTA": 78.254, "IDSW": 141, "IDF1": 81.106, "IDs": 372}

Conslusion