Fig. 1. Algorithm flow

Fig. 2. Coding process of L-Census

Fig. 3. Failure diagram in the boundary area of the object, in which the correct matching point of p is q, but the corresponding mismatching point is e obtained by Eq. (3). (a) Part of left image of Teddy; (b) part of right image of Teddy

Fig. 4. Illustration of the effectiveness of the adaptive weighted combining strategy of bidirectional gradient cost

Fig. 5. Disparity maps obtained by different horizontal and vertical gradient costs combining methods. (a) Original image; (b) real disparity map; (c) disparity map obtained using Eq. (3); (d) disparity map obtained using ABiGrad

Fig. 6. Construction of adaptive cross window. (a) Cross arm; (b) adaptive support area

Fig. 7. Area to verify the "candidate disparities" idea

Fig. 8. Flow chart of disparity calculation

Fig. 9. Pseudo code for disparity calculation

Fig. 10. Effect of parameters changing on error rate. (a) Variation of error rate with λ_Grad; (b) variation of error rate with λ_Census; (c) variation of error rate with M; (d) variation of error rate with τ_c; (e) variation of error rate with τ_d

Fig. 11. Disparity maps obtained by different cost calculation methods. (a) Reference image; (b) real disparity map; (c) AD-Cen; (d) AD-Grad; (e) LCen-ABiGrad

Fig. 12. Advantages of DC over WTA in the repeated texture area (straight frame), weak texture area (dotted frame) and untextured area (double straight frame). (a) Reference image; (b) disparity map obtained by WTA; (c) disparity map obtained by DC; (d) marked disparity map

Fig. 13. Results of our algorithm on standard stereo image pairs. (a) Reference image; (b) real disparity map; (c) disparity map generated by our algorithm; (d) mismatching map